Research on intelligent pharmacy unmanned delivery vehicle path planning and drug recognition technology based on JPS and RRT algorithms Authors: Hengzhe Yue, Kun Cao, Zhaokun Yang, Quntao An, Lizhi Sun Published: 2025 ABSTRACT This paper is set in the context of an intelligent pharmacy working environment, where we optimize the conventional path planning algorithms for existing drug-dispensing robots. By integrating the Jump Point Search algorithm with the Rapidly-exploring Random Tree Star algorithm, we enhance the accuracy of robot navigation and positioning.Utilizing CNN to fit the map images and setting thresholds to eliminate errors, thereby improving the accuracy and robustness of the map; optimizing the path algorithm to prevent local optima and deadlocks. We introduce a contour-based drug recognition method that uses the SAD algorithm for accurate identification of drug types and positions, ensuring high precision and efficiency during delivery. This ensures high precision and efficiency in drug recognition during the delivery process. This method improves planning time by 82% over A*, enhances path quality by 25% over Dijkstra, and reduces CPU usage by 68% compared with RRT*, while maintaining strong adaptability in dynamic environments. For drug recognition, the improved SAD algorithm lowers the mismatch rate from 27.33% (SAD) and 25.92% (SSD) to 15.12% on a 500-image dataset collected under mixed indoor lighting, achieving a 12.21% absolute gain and reducing recognition time by about 7 seconds. These values were measured on a 500-image pharmacy dataset collected under mixed indoor lighting. Introduction The ongoing intelligent transformation of logistics and supply chains has created a strong demand for autonomous mobile robots1–3. A particularly promising application is within the modern intelligent pharmacy, where unmanned delivery vehicles are increasingly deployed to handle critical tasks such as transporting medicines and organizing goods4,5. In such environments, efficient and reliable drug delivery is paramount and hinges on precise robotic path planning and accurate drug identification6. In the practical operational context of intelligent pharmacies, path planning for unmanned delivery vehicles confronts distinct and stringent challenges, rendering many general-purpose algorithms hardly applicable directly. First, pharmacy layouts typically encompass regular shelf aisles (highly structured) yet also involve low-frequency, poorly predictable dynamic activities from medical staff and mobile devices.This coexistence of environmental structuredness and dynamism causes pure static planners to be highly prone to path failure when encountering dynamic obstacles. Conversely, pure dynamic planners (e.g. RRT*) exhibit low efficiency in random sampling within wide, regular aisles, with planning times as long as 2.3 seconds, failing to meet real-time requirements. Second, to control costs and power consumption, systems are usually deployed on embedded platforms like Raspberry Pi, necessitating algorithms to operate stably within limited CPU and memory resources;complex algorithms or frequent global replanning can rapidly exhaust computing resources. Furthermore, there exists a sharp trade-off between path quality and safety: paths must not only be near-optimal in length but also sufficiently smooth to ensure stable vehicle operation, while maintaining safe distances from both static shelves and dynamic obstacles at all times–a contradiction that is particularly pronounced in narrow aisles.Thus, the core problem addressed in this study is: How to design a path planning method for pharmacy unmanned delivery vehicles running on low-power embedded platforms, enabling them to adaptively integrate the efficiency of global planning with the flexibility of local planning in complex environments where structure and dynamism coexist, thereby achieving a comprehensive balance of short planning time, high path quality, low resource consumption, and strong anti-interference capability. This project combines the Jump Point Search(JPS) algorithm7 for global static path planning and the Rapidly-exploring Random Tree Star(RRT*) algorithm8 for local dynamic adjustments, dynamically switching between them based on the environment. The switching is triggered primarily by real-time sensor data: when a dynamic obstacle is detected within a 2-meter radius of the vehicle’s current path segment, or when the local obstacle density exceeds 30%, the system shifts from JPS guided global planning to RRT* based local optimization. This integration enhances the robot’s navigation and positioning accuracy while incorporating deep learning techniques to optimize delivery strategies and efficiency, thereby improving overall performance and avoiding issues like local path optimality and local latency. Furthermore, to enhance the delivery system’s accuracy, this project introduces drug recognition technology, employing the Sum of Absolute Differences (SAD) template matching algorithm based on drug contour recognition9. This ensures precision and efficiency in drug identification and placement during delivery, enabling more intelligent unmanned pharmacy management. The JPS-RRT* hybrid mechanism proposed in this study is novel in that it designs a bidirectional complementary interaction paradigm based on environmental features. Unlike the common simple concatenation approach where the global planning result is used as the initial seed for the local planner, this mechanism achieves deep collaboration through dynamic role switching: in structured environments, the JPS algorithm takes the lead in generating the global path skeleton and intelligently guides the sampling space of RRT* with its path points, significantly enhancing the search efficiency; in unstructured areas or when encountering dynamic obstacles, the RRT algorithm takes over the lead to optimize the local trajectory and avoid obstacles, and feeds back the optimized feasible path information to the JPS module for global path evaluation. This JPS provides intelligent guidance for RRT*, and RRT* provides feasibility complementation for JPS bidirectional interaction mechanism effectively overcomes the limitations of single algorithms or master-slave hybrid frameworks, achieving a better balance among planning efficiency, path quality, and environmental adaptability. The main contributions of this work are threefold. Firstly, we propose a hybrid JPS-RRT* path planning algorithm that demonstrates superior efficiency and adaptability in pharmacy-like environments. Secondly, we develop a contour-enhanced drug recognition method based on an improved SAD algorithm, achieving high speed and accuracy on low-power devices. Thirdly, we provide comprehensive experimental validation in both simulated and physical environments, verifying the practical effectiveness of the proposed system. Related work Path planning techniques Current path planning algorithms are primarily divided into global and local path planning categories10. Global path planning algorithms can be further classified into static-oriented and dynamic-oriented algorithms based on environmental states, or traditional and intelligent algorithms based on whether perception and decision-making are separated. Traditional path planning algorithms are mainly based on graph search and geometric models. The classic Dijkstra11 and A*12 algorithms achieve shortest path search by traversing nodes, but their computational complexity grows exponentially with environmental scale. Their efficiency is significantly influenced by heuristic function design. Although these algorithms have been studied for decades, their quantitative efficiency differences in complex scenarios require further investigation. The jump point search (JPS) algorithm improves search efficiency by pruning redundant nodes. However, these methods struggle with continuous spaces and unstructured environments7. Sampling-based algorithms like RRT* and Probabilistic Roadmap (PRM) demonstrate flexibility in high-dimensional spaces (e.g. robotic arm motion planning) through probabilistic modeling and path connection mechanisms, but their paths exhibit strong randomness and insufficient convergence speed8,13,14. Subsequent improved algorithms like RRT* approach optimal paths through progressive optimization strategies, but computational costs still limit their practical applications. In common static planning environments, Dijkstra and A* remain the primary algorithms. This project tests these traditional algorithms in both static and dynamic environments, collecting and comparing their efficiency. After testing both, we use an improved JPS-RRT* algorithm for testing in the same environment and compare its efficiency and accuracy. Dijkstra algorithm The traditional Dijkstra algorithm is a classic method for finding single-source shortest paths in weighted graphs, applicable to cases with non-negative edge weights. The specific steps of Dijkstra’s algorithm are: Initialization:Set the starting point’s distance to 0, indicating zero distance to itself.Initialize distances of all other nodes to infinity, representing their unreachability.Create a priority queue to store all nodes to be processed, sorted by current known shortest distances.Process the priority queue: While the queue is not empty, repeat:Extract the node with minimum distance.Traverse neighboring nodes.Perform relaxation operations.Mark processed nodes: Once a node is extracted and processed from the queue, its shortest path is determined and no further distance updates are performed.Termination condition: When the priority queue is empty, the algorithm terminates. If a shortest path exists, it is now determined.Key points:Greedy strategy: Always extends the node with the current known shortest path, ensuring local optimality leads to global optimality.Non-negative edge weights: The algorithm assumes non-negative weights; once a node is processed, no shorter path can be found through other routes, making it unsuitable for graphs with negative edge weights.Fig. 1Topological diagram of Dijkstra’s algorithm. Initialization:Set the starting point’s distance to 0, indicating zero distance to itself.Initialize distances of all other nodes to infinity, representing their unreachability.Create a priority queue to store all nodes to be processed, sorted by current known shortest distances. Initialization:Set the starting point’s distance to 0, indicating zero distance to itself.Initialize distances of all other nodes to infinity, representing their unreachability.Create a priority queue to store all nodes to be processed, sorted by current known shortest distances. Set the starting point’s distance to 0, indicating zero distance to itself. Set the starting point’s distance to 0, indicating zero distance to itself. Initialize distances of all other nodes to infinity, representing their unreachability. Initialize distances of all other nodes to infinity, representing their unreachability. Create a priority queue to store all nodes to be processed, sorted by current known shortest distances. Create a priority queue to store all nodes to be processed, sorted by current known shortest distances. Process the priority queue: While the queue is not empty, repeat:Extract the node with minimum distance.Traverse neighboring nodes.Perform relaxation operations. Process the priority queue: While the queue is not empty, repeat:Extract the node with minimum distance.Traverse neighboring nodes.Perform relaxation operations. Extract the node with minimum distance. Extract the node with minimum distance. Traverse neighboring nodes. Traverse neighboring nodes. Perform relaxation operations. Perform relaxation operations. Mark processed nodes: Once a node is extracted and processed from the queue, its shortest path is determined and no further distance updates are performed. Mark processed nodes: Once a node is extracted and processed from the queue, its shortest path is determined and no further distance updates are performed. Termination condition: When the priority queue is empty, the algorithm terminates. If a shortest path exists, it is now determined. Termination condition: When the priority queue is empty, the algorithm terminates. If a shortest path exists, it is now determined. Greedy strategy: Always extends the node with the current known shortest path, ensuring local optimality leads to global optimality. Greedy strategy: Always extends the node with the current known shortest path, ensuring local optimality leads to global optimality. Non-negative edge weights: The algorithm assumes non-negative weights; once a node is processed, no shorter path can be found through other routes, making it unsuitable for graphs with negative edge weights. Non-negative edge weights: The algorithm assumes non-negative weights; once a node is processed, no shorter path can be found through other routes, making it unsuitable for graphs with negative edge weights. Topological diagram of Dijkstra’s algorithm. Time complexity: With a priority queue (min-heap), the time complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O((V + E)\log V)$$\end{document}, where V is the number of nodes and E is the number of edges. Using a Fibonacci heap can optimize this to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(E + V\log V)$$\end{document}. However, it cannot handle negative edge weights, which would produce incorrect shortest paths. Additionally, its efficiency decreases with large-scale graphs, requiring optimized data structures or heuristic functions. As shown in Fig. 1, this is the operating schematic diagram of the Dijkstra’s algorithm. A* algorithm The A* algorithm is a heuristic search algorithm that combines the completeness of Dijkstra’s algorithm with the efficiency of greedy best-first search, quickly finding optimal paths from start to goal in weighted graphs. Its core idea is to guide the search direction using a heuristic function, reducing unnecessary node expansions and significantly improving efficiency15,16. The algorithm primarily relies on the following evaluation function17 for path selection and judgment:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} f(n) = g(n) + h(n) \end{aligned}$$\end{document}Key data structures:Open list: Stores nodes to be expanded, sorted by f(n) (typically implemented as a priority queue).Closed list: Stores processed nodes to avoid redundant calculations.Steps: Initialization: Add the starting point to the open list, set its g value to 0 and h value to the heuristic estimate, compute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f = g + h$$\end{document}, and set g values of other nodes to infinite .Process the open list in a loop:Select a node.Check if it’s the goal.Expand neighbors.Mark as processed.Termination condition: If the open list is empty and the goal hasn’t been reached, no path exists.The key mechanism lies in the design of the heuristic function, which must satisfy two properties: admissibility and consistency. Common heuristic functions include:Manhattan distance (suitable for grid movement in four directions)18: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = |x_n - x_{goal}| + |y_n - y_{goal}| \end{aligned}$$\end{document}Euclidean distance (suitable for continuous spaces): 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = \sqrt{(x_n - x_{goal})^2 + (y_n - y_{goal})^2} \end{aligned}$$\end{document}Diagonal distance (eight-direction movement): 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = \max (|x_n - x_{goal}|, |y_n - y_{goal}|) \end{aligned}$$\end{document}Fig. 2Path comparison between A* and Dijkstra algorithms. Open list: Stores nodes to be expanded, sorted by f(n) (typically implemented as a priority queue). Open list: Stores nodes to be expanded, sorted by f(n) (typically implemented as a priority queue). Closed list: Stores processed nodes to avoid redundant calculations. Closed list: Stores processed nodes to avoid redundant calculations. Initialization: Add the starting point to the open list, set its g value to 0 and h value to the heuristic estimate, compute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f = g + h$$\end{document}, and set g values of other nodes to infinite . Initialization: Add the starting point to the open list, set its g value to 0 and h value to the heuristic estimate, compute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f = g + h$$\end{document}, and set g values of other nodes to infinite . Process the open list in a loop:Select a node.Check if it’s the goal.Expand neighbors.Mark as processed. Process the open list in a loop:Select a node.Check if it’s the goal.Expand neighbors.Mark as processed. Select a node. Select a node. Check if it’s the goal. Check if it’s the goal. Expand neighbors. Expand neighbors. Mark as processed. Mark as processed. Termination condition: If the open list is empty and the goal hasn’t been reached, no path exists. Termination condition: If the open list is empty and the goal hasn’t been reached, no path exists. Manhattan distance (suitable for grid movement in four directions)18: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = |x_n - x_{goal}| + |y_n - y_{goal}| \end{aligned}$$\end{document} Manhattan distance (suitable for grid movement in four directions)18: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = |x_n - x_{goal}| + |y_n - y_{goal}| \end{aligned}$$\end{document} Euclidean distance (suitable for continuous spaces): 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = \sqrt{(x_n - x_{goal})^2 + (y_n - y_{goal})^2} \end{aligned}$$\end{document} Euclidean distance (suitable for continuous spaces): 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = \sqrt{(x_n - x_{goal})^2 + (y_n - y_{goal})^2} \end{aligned}$$\end{document} Diagonal distance (eight-direction movement): 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = \max (|x_n - x_{goal}|, |y_n - y_{goal}|) \end{aligned}$$\end{document} Diagonal distance (eight-direction movement): 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h(n) = \max (|x_n - x_{goal}|, |y_n - y_{goal}|) \end{aligned}$$\end{document} Path comparison between A* and Dijkstra algorithms. As shown in Fig. 2, this is the simulation roadmap of Dijkstra’s algorithm and the A* algorithm. D* algorithm The D* algorithm (Dynamic A) is a well-known incremental search algorithm designed for dynamic environments. Its core advantage lies in efficiently repairing paths when obstacle positions change, by incrementally updating costs and propagating these changes from the goal node backwards (reverse search), thus avoiding complete replanning from scratch. This makes D particularly suitable for scenarios with frequent, unpredictable obstacles. However, in the context of a structured pharmacy environment with predominantly static layouts and sparse, low-frequency dynamic obstacles (e.g. occasional staff movement), D’s general-purpose incremental repair mechanism may introduce unnecessary computational overhead. In contrast, the JPS-RRT* hybrid algorithm proposed in this work adopts a different strategy: it leverages the efficiency of JPS for global planning in the static parts of the map and strategically invokes RRT* only for local replanning around detected dynamic obstacles or in complex areas. This approach is specifically tailored to environments where the global map is largely static and known, allowing for more efficient use of computational resources on embedded platforms like the Raspberry Pi. Our experimental results in Sect. "Performance comparison" demonstrate that JPS-RRT* achieves a better balance between planning time, path quality, and adaptability to dynamic changes in such structured environments compared to what would be expected from a standard D* implementation.As shown in Fig. 3, this is the flowchart of the D* algorithm.Fig. 3Workflow of the D* algorithm. Workflow of the D* algorithm. Theoretical comparison Here are two tables respectively presenting the theoretical comparison of Dijkstra and A* algorithms as shown in Table 1, and the theoretical performance analysis of Dijkstra and A* algorithms as shown in Table 2.Table 1Theoretical comparison between Dijkstra and A* algorithms.MetricDijkstraA* (admissible heuristic)OptimalityGuaranteedGuaranteedSearch directionOmnidirectional expansionGoal-orientedMemory consumptionHigh (stores all nodes)Low (prioritizes key nodes)Table 2Theoretical performance analysis of Dijkstra and A* algorithms.MetricDijkstraA*OptimalityGuaranteedGuaranteed (admissible heuristic)Time complexity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O((V + E)\log V)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(b^d)$$\end{document}Space complexityO(V)O(V)Applicable spaceGlobal searchGoal-oriented search Table 1Theoretical comparison between Dijkstra and A* algorithms.MetricDijkstraA* (admissible heuristic)OptimalityGuaranteedGuaranteedSearch directionOmnidirectional expansionGoal-orientedMemory consumptionHigh (stores all nodes)Low (prioritizes key nodes) Theoretical comparison between Dijkstra and A* algorithms. Table 2Theoretical performance analysis of Dijkstra and A* algorithms.MetricDijkstraA*OptimalityGuaranteedGuaranteed (admissible heuristic)Time complexity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O((V + E)\log V)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(b^d)$$\end{document}Space complexityO(V)O(V)Applicable spaceGlobal searchGoal-oriented search Theoretical performance analysis of Dijkstra and A* algorithms. To validate the theoretical comparisons above and evaluate the performance of these algorithms in our specific pharmacy environment, we conducted empirical tests on an embedded platform (Raspberry Pi 4B). The key findings, summarized from Sect. “Test results for Dijkstra and A* algorithms", indicate that the A algorithm significantly outperformed Dijkstra in planning time (132±9 ms vs. 572±23 ms) and resource consumption (CPU load: 53±2% vs. 78±3%), making it more suitable for high-frequency planning tasks. However, both algorithms exhibited limitations in dynamic environments, such as positioning errors and motion deadlocks when encountering dynamic obstacles. These limitations motivated the development of our improved JPS-RRT* hybrid algorithm, detailed in Sect. “New fusion algorithm". For comprehensive experimental data and analysis, readers are referred to Sect. “Test results for Dijkstra and A* algorithms" and Table 4. Experimental design for Dijkstra and A* algorithms Test environment Simulation environment:Pharmacy warehouse scenario (static obstacles)Dynamic scenario (random pedestrian movement)Physical environment:Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with fixed obstacles at 20%-50% density)Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with 2 dynamic obstacles)The pharmacy warehouse scenario was modeled as a grid map with a resolution of 5 cm per cell, translating to a grid size of 160x120 cells for the 8m x 6m physical area.Obstacle arrangement: For the test site, objects with dimensions of 20 cm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 30 cm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 5 cm were selected as obstacles. A total of three static obstacles were randomly placed to simulate the position of the medicine cabinet in a pharmacy. Additionally, two dynamic obstacles were set up to move at a constant speed back and forth within a fixed area, to simulate the movement of pharmacy staff and the mutual avoidance when multiple unmanned vehicles are working together.Evaluation metrics:As shown in Table 3, what is represented in this table are the evaluation indicators of the algorithm performance.Table 3Evaluation metrics for algorithm performance.CategoryMetricMeasurement methodAlgorithm efficiencyPlanning timeRaspberry Pi system clock timingResource usageCPU utilization, memory peakCommand sampling (1Hz)Path qualityLength, smoothnessSimulation: theoretical vs. actual valuesTracking performanceLateral deviationPhysical: Laser positioning data vs. desired path Simulation environment:Pharmacy warehouse scenario (static obstacles)Dynamic scenario (random pedestrian movement) Simulation environment:Pharmacy warehouse scenario (static obstacles)Dynamic scenario (random pedestrian movement) Pharmacy warehouse scenario (static obstacles) Pharmacy warehouse scenario (static obstacles) Dynamic scenario (random pedestrian movement) Dynamic scenario (random pedestrian movement) Physical environment:Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with fixed obstacles at 20%-50% density)Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with 2 dynamic obstacles)The pharmacy warehouse scenario was modeled as a grid map with a resolution of 5 cm per cell, translating to a grid size of 160x120 cells for the 8m x 6m physical area. Physical environment:Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with fixed obstacles at 20%-50% density)Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with 2 dynamic obstacles)The pharmacy warehouse scenario was modeled as a grid map with a resolution of 5 cm per cell, translating to a grid size of 160x120 cells for the 8m x 6m physical area. Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with fixed obstacles at 20%-50% density) Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with fixed obstacles at 20%-50% density) Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with 2 dynamic obstacles) Indoor test field (8m\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}6m, with 2 dynamic obstacles) The pharmacy warehouse scenario was modeled as a grid map with a resolution of 5 cm per cell, translating to a grid size of 160x120 cells for the 8m x 6m physical area. The pharmacy warehouse scenario was modeled as a grid map with a resolution of 5 cm per cell, translating to a grid size of 160x120 cells for the 8m x 6m physical area. Obstacle arrangement: For the test site, objects with dimensions of 20 cm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 30 cm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 5 cm were selected as obstacles. A total of three static obstacles were randomly placed to simulate the position of the medicine cabinet in a pharmacy. Additionally, two dynamic obstacles were set up to move at a constant speed back and forth within a fixed area, to simulate the movement of pharmacy staff and the mutual avoidance when multiple unmanned vehicles are working together. Obstacle arrangement: For the test site, objects with dimensions of 20 cm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 30 cm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 5 cm were selected as obstacles. A total of three static obstacles were randomly placed to simulate the position of the medicine cabinet in a pharmacy. Additionally, two dynamic obstacles were set up to move at a constant speed back and forth within a fixed area, to simulate the movement of pharmacy staff and the mutual avoidance when multiple unmanned vehicles are working together. Evaluation metrics: Evaluation metrics: Table 3Evaluation metrics for algorithm performance.CategoryMetricMeasurement methodAlgorithm efficiencyPlanning timeRaspberry Pi system clock timingResource usageCPU utilization, memory peakCommand sampling (1Hz)Path qualityLength, smoothnessSimulation: theoretical vs. actual valuesTracking performanceLateral deviationPhysical: Laser positioning data vs. desired path Evaluation metrics for algorithm performance. Testing process for Dijkstra and A* Algorithms By establishing fixed motion paths, we tested the operational efficiency of both algorithms, including their turning frequency at corners. Simultaneously, we monitored various loads on the Raspberry Pi. Finally, we compared whether both algorithms could accurately reach specified target points under identical positioning conditions. The experimental scenario is shown in the SLAM mapping in Fig. 4.Fig. 4Experimental scenario. Experimental scenario. Simultaneous Localization and Mapping (SLAM) enables autonomous vehicles to operate in unknown environments, building environmental maps while self-localizing. Autonomous vehicles estimate their positions using multiple sensors while updating their positions and orientations based on grid maps19. Common SLAM mapping algorithms include Cartographer and Gmapping. However, Cartographer is typically used for large-scale environment modeling and requires high computational resources. Given our project’s application in smart pharmacy environments, the Gmapping algorithm is more suitable. Gmapping relies on lidar and odometry to obtain environmental pose information, optimizes particle distribution, and introduces an adaptive resampling mechanism, improving algorithm efficiency20. Therefore, we ultimately selected Gmapping for mapping. However, single mapping sessions often produce maps with excessive noise and inaccuracies in certain obstacles and corners. To address this, we introduced deep learning to automatically match and correct the generated maps. Through the vehicle’s dynamic tracking capability, we achieved an artificial teaching effect, enabling autonomous tracking and map building. Subsequently, we used deep learning to match overlapping sections from multiple recorded map datasets as the actual modeling map. After validation, the accuracy of maps trained multiple times reached 95%-99%. After establishing the standard map, we tested positioning errors, planning efficiency, and Raspberry Pi utilization under different algorithms. The restored and unrestored map models are respectively presented in Figs. 5 and 6.Fig. 5After optimization.Fig. 6Before optimization. After optimization. Before optimization. Test results for Dijkstra and A* algorithms Table 4Algorithm performance on Raspberry Pi 4B.AlgorithmPlanning time (ms)CPU load (%)Memory peak (MB)Dijkstra \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$572 \pm 23$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$78 \pm 3$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$215 \pm 8$$\end{document} A* \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$132 \pm 9$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$53 \pm 2$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$127 \pm 5$$\end{document} Table 4Algorithm performance on Raspberry Pi 4B.AlgorithmPlanning time (ms)CPU load (%)Memory peak (MB)Dijkstra \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$572 \pm 23$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$78 \pm 3$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$215 \pm 8$$\end{document} A* \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$132 \pm 9$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$53 \pm 2$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$127 \pm 5$$\end{document} Algorithm performance on Raspberry Pi 4B. The results demonstrate that A* significantly reduces computational resource demands, making it suitable for high-frequency repetitive planning tasks. Comparing actual map positions with their locations revealed certain deviations, including positioning errors and wheel slippage causing odometry inaccuracies. Specific errors were:A* algorithm: Average error of 0.14m (maximum 0.23m) due to larger path turning angles.Dijkstra algorithm: Average deviation of 0.09m (maximum 0.15m) with smoother paths.Both algorithms exhibit distinct issues: A* offers faster response and decision-making speeds than Dijkstra but with larger errors; conversely, Dijkstra has smaller errors but slower planning speeds and higher resource consumption. A* algorithm: Average error of 0.14m (maximum 0.23m) due to larger path turning angles. A* algorithm: Average error of 0.14m (maximum 0.23m) due to larger path turning angles. Dijkstra algorithm: Average deviation of 0.09m (maximum 0.15m) with smoother paths. Dijkstra algorithm: Average deviation of 0.09m (maximum 0.15m) with smoother paths. Additionally, both perform poorly with dynamic obstacles, experiencing positioning deviations, coordinate errors, and motion deadlocks. To address these issues, we ultimately selected an improved combination of JPS (an enhanced A* algorithm) and RRT* as our path planning algorithm. This combination meets rapid planning requirements while reducing errors, with RRT* compensating for A*’s poor performance in dynamic environments.As shown in Table 4, it is the algorithmic performance of the Raspberry PI 4B. High-precision map verification using absolute trajectory error (ATE) To address the critical dependency of the JPS algorithm on map accuracy, we implemented a rigorous quantitative verification method based on Absolute Trajectory Error (ATE).This approach moves beyond qualitative assessment and provides a concrete, measurable criterion for map quality, ensuring the reliability of the global path planner. The verification process consisted of three key steps: Ground-truth trajectory acquisition: During the SLAM data collection phase, we established a high-precision ground-truth reference system.This system provides centimeter-level positioning accuracy, serving as a reliable benchmark against which the SLAM-estimated trajectory could be compared.Trajectory estimation and synchronization: The Gmapping SLAM algorithm was used to process the laser and odometry data, generating an estimated trajectory of the robot and the corresponding grid map.The estimated trajectory from SLAM was then precisely synchronized with the ground-truth trajectory using timestamps.Calculation and analysis of ATE: The core of verification is the calculation of ATE. For each synchronous pose in the trajectory, we calculate its absolute accuracy and then quantify the overall map’s accuracy through the root mean square error (RMSE) of these individual errors. We utilized the widely adopted EVO evaluation toolkit to automatically perform this calculation and generate statistical results.The resulting ATE RMSE for our final map was 0.038 meters.This metric objectively confirms that the geometric distortion in the constructed map is minimal.A map with such high global consistency is fully capable of supporting the JPS algorithm, as the relative positions of obstacles and free spaces are accurately preserved.This method is a commonly used one in the current market. Its specific operation process needs to be carried out based on the actual situation. Relevant research can be referred to for deployment. Ground-truth trajectory acquisition: During the SLAM data collection phase, we established a high-precision ground-truth reference system.This system provides centimeter-level positioning accuracy, serving as a reliable benchmark against which the SLAM-estimated trajectory could be compared. Ground-truth trajectory acquisition: During the SLAM data collection phase, we established a high-precision ground-truth reference system.This system provides centimeter-level positioning accuracy, serving as a reliable benchmark against which the SLAM-estimated trajectory could be compared. Trajectory estimation and synchronization: The Gmapping SLAM algorithm was used to process the laser and odometry data, generating an estimated trajectory of the robot and the corresponding grid map.The estimated trajectory from SLAM was then precisely synchronized with the ground-truth trajectory using timestamps. Trajectory estimation and synchronization: The Gmapping SLAM algorithm was used to process the laser and odometry data, generating an estimated trajectory of the robot and the corresponding grid map.The estimated trajectory from SLAM was then precisely synchronized with the ground-truth trajectory using timestamps. Calculation and analysis of ATE: The core of verification is the calculation of ATE. For each synchronous pose in the trajectory, we calculate its absolute accuracy and then quantify the overall map’s accuracy through the root mean square error (RMSE) of these individual errors. We utilized the widely adopted EVO evaluation toolkit to automatically perform this calculation and generate statistical results. Calculation and analysis of ATE: The core of verification is the calculation of ATE. For each synchronous pose in the trajectory, we calculate its absolute accuracy and then quantify the overall map’s accuracy through the root mean square error (RMSE) of these individual errors. We utilized the widely adopted EVO evaluation toolkit to automatically perform this calculation and generate statistical results. Drug recognition techniques Evolution and limitations of traditional image processing methods Pharmaceutical image recognition technology serves as a critical component in intelligent pharmaceutical systems, demonstrating a clear technological evolution from traditional image processing methods to deep learning approaches. Early research primarily focused on conventional image processing techniques that relied on handcrafted feature extractors. Among these, template matching technology represents one of the most significant approaches, operating on the fundamental principle of calculating similarity between target images and predefined templates for identification. While this study employs the Sum of Absolute Differences (SAD) algorithm, the Normalized Cross-Correlation (NCC) algorithm has been widely adopted in drug recognition scenarios due to its robustness to illumination variations21. Concurrently, algorithms such as Scale-Invariant Feature Transform (SIFT)22 and Speeded-Up Robust Features (SURF)23 have been utilized to handle tablet shape and marking matching tasks by extracting local keypoint features, partially addressing scale and rotation variations. However, these traditional methods generally suffer from inherent limitations including sensitivity to image quality and computational complexity that scales linearly with template quantity, presenting significant challenges in real-time embedded applications. Breakthroughs and challenges in deep learning technologies Parallel to traditional methods, deep learning-based pharmaceutical recognition technologies have emerged with breakthrough improvements in identification accuracy under ideal conditions through end-to-end feature learning mechanisms. Two-stage detection algorithms (e.g. the R-CNN series)24 demonstrate exceptional precision in offline quality inspection scenarios through their cascaded pipeline of region proposal generation and refined classification, though their complex computational workflows struggle to meet real-time requirements. In contrast, single-stage detection algorithms (such as YOLO and SSD)25,26 frame object detection as a regression problem, achieving superior balance between speed and accuracy, making them preferred solutions for real-time sorting systems. Notably, the recent emergence of lightweight network architectures (e.g. combinations of MobileNet and YOLO)27 has created new possibilities for embedded device deployment. Nevertheless, the advantages of deep learning methods remain highly dependent on large-scale annotated datasets, while maintaining high-frame-rate inference under extremely resource-constrained environments continues to pose practical challenges. Innovative positioning and engineering trade-offs of this study Within this technological spectrum, the improved SAD method proposed in this study establishes a unique positioning: it does not pursue ultimate accuracy in general scenarios but rather offers a specialized optimization solution tailored for structured environments such as intelligent pharmacies. Compared to traditional template matching, this approach addresses computational efficiency bottlenecks through a contour pre-classification mechanism that constructs a hierarchical template library. Relative to deep learning methods, its core advantages manifest in no training requirements, minimal resource consumption, and strong interpretability. This design philosophy enables the algorithm to efficiently collaborate with path planning modules on unified embedded platforms, ultimately achieving system-level performance optimization that embodies the engineering trade-off wisdom oriented toward practical application scenarios. Moreover, the proposed improved SAD method fills a practical research gap that is often neglected in deep learning–oriented drug recognition studies. While existing deep learning approaches achieve high accuracy, they implicitly assume the availability of large annotated datasets, sufficient GPU resources, and stable high-resolution imaging. However, intelligent pharmacy systems frequently operate under small-sample conditions, tight computational budgets, and strict real-time constraints, where deploying or updating deep models becomes impractical. Current literature offers limited solutions for these resource-constrained and highly structured environments. By integrating contour-based pre-classification with a hierarchical template library, the improved SAD method provides a lightweight yet reliable alternative specifically optimized for these scenarios. Thus, instead of competing with deep learning approaches in general-purpose recognition tasks, this study complements them by addressing their deployment limitations, thereby filling an important engineering-focused gap in pharmaceutical automation research. New fusion algorithm JPS-RRT* algorithm The Jump Point Search (JPS) algorithm is a path planning algorithm optimized for grid maps, designed to significantly reduce the number of node expansions in A* algorithms. Without environmental preprocessing, it achieves notable acceleration effects. Its core idea involves pruning symmetric paths and directly skipping redundant nodes, retaining only key nodes (jump points) for expansion. This makes it suitable for structured grid environments (e.g. game maps, warehouse layouts)28.Fig. 7Path diagram of JPS algorithm. Path diagram of JPS algorithm. As shown in Fig. 7, it is the simulation route when JPS is conducting path planning. In environments with preprocessing, the JPS algorithm can be further optimized into JPS+, achieving even more significant acceleration29. The key lies in two aspects:Forced neighbors: When a node moves in a certain direction and must bypass an obstacle, the obstacle’s adjacent nodes are called forced neighbors (requiring reachability checks).Jump points: Nodes satisfying any of the following conditions: Are start or goal points.Have forced neighbors.Are endpoints of “straight-line jumps” toward the goal. Forced neighbors: When a node moves in a certain direction and must bypass an obstacle, the obstacle’s adjacent nodes are called forced neighbors (requiring reachability checks). Forced neighbors: When a node moves in a certain direction and must bypass an obstacle, the obstacle’s adjacent nodes are called forced neighbors (requiring reachability checks). Jump points: Nodes satisfying any of the following conditions: Jump points: Nodes satisfying any of the following conditions: Are start or goal points. Are start or goal points. Have forced neighbors. Have forced neighbors. Are endpoints of “straight-line jumps” toward the goal. Are endpoints of “straight-line jumps” toward the goal. Algorithm 1Jump point search jump function. Fig. 8Path diagram of RRT algorithm. Jump point search jump function. Path diagram of RRT algorithm. Through JPS’s jump point symmetry pruning, the number of nodes is reduced from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(n^2)$$\end{document} in A* to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(n \log n)$$\end{document}.The Algorithm 1 here is the pseudo-code part of the JPS algorithm. The RRT* algorithm is a sampling-based path planning algorithm designed to solve optimal path search problems in high-dimensional complex environments. Building upon the basic RRT algorithm, it introduces progressive optimization mechanisms that continuously adjust the tree structure to approach theoretically optimal paths.As shown in Fig. 8, it is the path simulation diagram of the RRT algorithm. The algorithm flow mainly consists of: Initialize parameters:Input: Start point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {start}}$$\end{document}, goal region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {goal}}$$\end{document}, configuration space C, collision detection function.Output: Tree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T = (V, E)$$\end{document}, where V is the node set and E is the edge set.Initialize tree: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V \leftarrow \{q_{\text {start}}\}, E \leftarrow \emptyset$$\end{document}.Randomly generate a point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} in configuration space C (goal-biased sampling: directly sample the goal with a certain probability to accelerate convergence).Find the nearest node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {near}}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} in tree T: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{\text {near}} = \arg \min _{q \in V} \Vert q - q_{\text {rand}}\Vert \end{aligned}$$\end{document}Move from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {near}}$$\end{document} toward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} with step size n to generate new node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}.Check if path segment \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L(q_{\text {near}}, q_{\text {new}})$$\end{document} is collision-free; if not, discard the node and resample.In tree T, find all nodes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {near}}$$\end{document} within radius r centered at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}.Select optimal parent node: From \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {near}} \cup \{q_{\text {near}}\}$$\end{document}, choose parent node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {parent}}$$\end{document} that minimizes the cost from start to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{\text {parent}} = \arg \min _{q \in Q_{\text {near}} \cup \{q_{\text {near}}\}} (\text {Cost}(q) + \Vert q - q_{\text {new}}\Vert ) \end{aligned}$$\end{document} Add \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document} to V and edge \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(q_{\text {parent}}, q_{\text {new}})$$\end{document} to E.Rewire: For each neighboring node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q \in Q_{\text {near}}$$\end{document}, check if a lower cost path exists through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {if } \text {Cost}(q_{\text {new}}) + \Vert q_{\text {new}} - q\Vert < \text {Cost}(q) \text { then} \end{aligned}$$\end{document}Key aspects of this algorithm include:Neighboring node radius selection: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r = \gamma \left( \frac{\log N}{N} \right) ^{1/d}, \end{aligned}$$\end{document} where N is the current number of tree nodes, d is the configuration space dimension, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document} is a tuning parameter (typically 2-3).Path cost: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Cost}(q) = \sum _{i=1}^{k} \Vert q_i - q_{i-1}\Vert \quad (q_0 = q_s). \end{aligned}$$\end{document}Although JPS and RRT* show significant efficiency improvements and reduced resource consumption compared to A* and Dijkstra, they still have respective limitations:RRT* limitations: Relies on random sampling and progressive optimization, requiring numerous iterations to approach optimal paths (time complexity: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(n \log n)$$\end{document}).In regular grids or channelized environments, cannot leverage environmental structures to accelerate searches, leading to redundant sampling.Requires storing complete tree structures, with node counts growing exponentially with environmental complexity.JPS limitations: Cannot handle continuous spaces or unstructured environments (e.g. complex terrains, high-dimensional configuration spaces).Paths depend on static maps; dynamic obstacles require complete jump point researching.Paths consist of line segments requiring post-processing smoothing, potentially violating kinematic constraints (e.g. curvature continuity).By combining these two algorithms, we can optimize certain issues, thereby improving efficiency and accuracy. First, we address RRT*’s inefficiency in structured environments by using JPS jump points to quickly generate initial paths in regular regions, reducing RRT*’s global sampling needs30. Simultaneously, we mitigate JPS’s failure in unstructured environments by switching to RRT* optimization in complex scenarios, avoiding JPS path infeasibility. For dynamic obstacle handling, both algorithms collaborate: JPS provides a global static path framework, while RRT* performs local real-time adjustments upon detecting dynamic obstacles. This approach meets dynamic environment requirements while minimizing replanning scope. Initialize parameters:Input: Start point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {start}}$$\end{document}, goal region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {goal}}$$\end{document}, configuration space C, collision detection function.Output: Tree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T = (V, E)$$\end{document}, where V is the node set and E is the edge set.Initialize tree: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V \leftarrow \{q_{\text {start}}\}, E \leftarrow \emptyset$$\end{document}. Initialize parameters:Input: Start point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {start}}$$\end{document}, goal region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {goal}}$$\end{document}, configuration space C, collision detection function.Output: Tree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T = (V, E)$$\end{document}, where V is the node set and E is the edge set.Initialize tree: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V \leftarrow \{q_{\text {start}}\}, E \leftarrow \emptyset$$\end{document}. Input: Start point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {start}}$$\end{document}, goal region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {goal}}$$\end{document}, configuration space C, collision detection function. Input: Start point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {start}}$$\end{document}, goal region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {goal}}$$\end{document}, configuration space C, collision detection function. Output: Tree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T = (V, E)$$\end{document}, where V is the node set and E is the edge set. Output: Tree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T = (V, E)$$\end{document}, where V is the node set and E is the edge set. Initialize tree: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V \leftarrow \{q_{\text {start}}\}, E \leftarrow \emptyset$$\end{document}. Initialize tree: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V \leftarrow \{q_{\text {start}}\}, E \leftarrow \emptyset$$\end{document}. Randomly generate a point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} in configuration space C (goal-biased sampling: directly sample the goal with a certain probability to accelerate convergence). Randomly generate a point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} in configuration space C (goal-biased sampling: directly sample the goal with a certain probability to accelerate convergence). Find the nearest node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {near}}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} in tree T: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{\text {near}} = \arg \min _{q \in V} \Vert q - q_{\text {rand}}\Vert \end{aligned}$$\end{document} Find the nearest node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {near}}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} in tree T: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{\text {near}} = \arg \min _{q \in V} \Vert q - q_{\text {rand}}\Vert \end{aligned}$$\end{document} Move from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {near}}$$\end{document} toward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} with step size n to generate new node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}. Move from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {near}}$$\end{document} toward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {rand}}$$\end{document} with step size n to generate new node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}. Check if path segment \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L(q_{\text {near}}, q_{\text {new}})$$\end{document} is collision-free; if not, discard the node and resample. Check if path segment \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L(q_{\text {near}}, q_{\text {new}})$$\end{document} is collision-free; if not, discard the node and resample. In tree T, find all nodes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {near}}$$\end{document} within radius r centered at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}. In tree T, find all nodes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {near}}$$\end{document} within radius r centered at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}. Select optimal parent node: From \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {near}} \cup \{q_{\text {near}}\}$$\end{document}, choose parent node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {parent}}$$\end{document} that minimizes the cost from start to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{\text {parent}} = \arg \min _{q \in Q_{\text {near}} \cup \{q_{\text {near}}\}} (\text {Cost}(q) + \Vert q - q_{\text {new}}\Vert ) \end{aligned}$$\end{document} Add \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document} to V and edge \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(q_{\text {parent}}, q_{\text {new}})$$\end{document} to E. Select optimal parent node: From \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{\text {near}} \cup \{q_{\text {near}}\}$$\end{document}, choose parent node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {parent}}$$\end{document} that minimizes the cost from start to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{\text {parent}} = \arg \min _{q \in Q_{\text {near}} \cup \{q_{\text {near}}\}} (\text {Cost}(q) + \Vert q - q_{\text {new}}\Vert ) \end{aligned}$$\end{document} Add \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document} to V and edge \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(q_{\text {parent}}, q_{\text {new}})$$\end{document} to E. Rewire: For each neighboring node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q \in Q_{\text {near}}$$\end{document}, check if a lower cost path exists through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {if } \text {Cost}(q_{\text {new}}) + \Vert q_{\text {new}} - q\Vert < \text {Cost}(q) \text { then} \end{aligned}$$\end{document} Rewire: For each neighboring node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q \in Q_{\text {near}}$$\end{document}, check if a lower cost path exists through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{\text {new}}$$\end{document}: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {if } \text {Cost}(q_{\text {new}}) + \Vert q_{\text {new}} - q\Vert < \text {Cost}(q) \text { then} \end{aligned}$$\end{document} Neighboring node radius selection: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r = \gamma \left( \frac{\log N}{N} \right) ^{1/d}, \end{aligned}$$\end{document} where N is the current number of tree nodes, d is the configuration space dimension, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document} is a tuning parameter (typically 2-3). Neighboring node radius selection: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r = \gamma \left( \frac{\log N}{N} \right) ^{1/d}, \end{aligned}$$\end{document} where N is the current number of tree nodes, d is the configuration space dimension, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document} is a tuning parameter (typically 2-3). Path cost: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Cost}(q) = \sum _{i=1}^{k} \Vert q_i - q_{i-1}\Vert \quad (q_0 = q_s). \end{aligned}$$\end{document} Path cost: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Cost}(q) = \sum _{i=1}^{k} \Vert q_i - q_{i-1}\Vert \quad (q_0 = q_s). \end{aligned}$$\end{document} RRT* limitations: RRT* limitations: Relies on random sampling and progressive optimization, requiring numerous iterations to approach optimal paths (time complexity: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(n \log n)$$\end{document}). Relies on random sampling and progressive optimization, requiring numerous iterations to approach optimal paths (time complexity: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(n \log n)$$\end{document}). In regular grids or channelized environments, cannot leverage environmental structures to accelerate searches, leading to redundant sampling. In regular grids or channelized environments, cannot leverage environmental structures to accelerate searches, leading to redundant sampling. Requires storing complete tree structures, with node counts growing exponentially with environmental complexity. Requires storing complete tree structures, with node counts growing exponentially with environmental complexity. JPS limitations: JPS limitations: Cannot handle continuous spaces or unstructured environments (e.g. complex terrains, high-dimensional configuration spaces). Cannot handle continuous spaces or unstructured environments (e.g. complex terrains, high-dimensional configuration spaces). Paths depend on static maps; dynamic obstacles require complete jump point researching. Paths depend on static maps; dynamic obstacles require complete jump point researching. Paths consist of line segments requiring post-processing smoothing, potentially violating kinematic constraints (e.g. curvature continuity). Paths consist of line segments requiring post-processing smoothing, potentially violating kinematic constraints (e.g. curvature continuity). JPS-RRT* algorithm operation Interaction process Global path generation (JPS-dominant):Input: Grid map (from SLAM), start and goal points.Process: JPS quickly prunes symmetric paths in the grid map, generating an initial path composed of jump points.The JPS path is segmented by curvature or obstacle density, with regions needing optimization marked (e.g. narrow passages, complex turns).Output: Line segments and optimization requirement markers.Local optimization (RRT*-dominant):Input: JPS path segments, continuous space map, optimization markers.Process: Map JPS segments to continuous space, generating adjustable-width safety corridors:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W = W_\text {base} + \alpha \cdot \text {obstacle}_\text {density} \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}}$$\end{document} is the base width and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} is a tuning factor. The parameters in Eq. (10) were selected based on a combination of physical constraints and systematic optimization objectives. The base width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}} = 1.0\,\textrm{m}$$\end{document} is derived from the vehicle’s physical width (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {vehicle}} \approx 0.5\,\textrm{m}$$\end{document}) plus a necessary safety margin on both sides. This margin accounts for positioning errors (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 0.1\,\textrm{m}$$\end{document}, see Sect. Test results for Dijkstra and A* algorithms) and tracking control deviations, and is consistent with the common practice of setting an inflation distance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2 \times 0.25\,\textrm{m}$$\end{document}) around the robot. Formally, 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W_{\text {base}} \ge W_{\text {vehicle}} + 2 \times (\text {Positioning Error} + \text {Control Buffer}), \end{aligned}$$\end{document} ensuring a minimum clearance for safe passage. The tuning factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was determined by formulating the corridor width selection as an optimization problem with two competing objectives: (1) maximizing path feasibility (favored by a wider corridor), and (2) minimizing computational load (favored by a narrower corridor). A wider corridor increases the probability that RRT* can find a feasible path quickly but also enlarges the sampling space, potentially increasing planning time. We quantified these effects through a grid search over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \in [0.2, 1.0]$$\end{document} in our simulation environment. The primary optimization criteria were to achieve zero collisions while minimizing the planning time. The value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was found to be the Pareto-optimal point that best balanced these objectives for our specific pharmacy environment, leading to a significant reduction in planning time compared to a fixed-width corridor (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0$$\end{document}), without incurring additional collision risk. This approach provides a more principled rationale beyond pure empirical observation. For the value range of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}, we generally refer to the following range and then make specific adjustments based on the actual situation on site.Indoor/Narrow passage 0.5-1.0.Outdoor/low-density scenes: 0.2-0.5.Dynamic scenes (such as pedestrians) 0.8-1.2.(b)Sample with high probability (e.g. 70%) within safety corridors to guide RRT* tree growth along JPS paths. Sampling probability distribution: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_{\text {bias}} = 0.7 \text { (corridor)}, \quad p_{\text {uniform}} = 0.3 \text { (full space)} \end{aligned}$$\end{document} A sampling rate exceeding 70% can accelerate convergence, but it is prone to getting stuck in local optima or dead zones; a rate lower than 70% leads to scattered sampling, increased computational load, and slower convergence. Therefore, a sampling rate of 70% is selected to balance smooth path planning and response to sudden obstacles. The width of the safety corridor needs to be designed based on the actual scenario; it should not be too narrow, as it is difficult to plan, and it should not be too wide, as it increases the computational load. Equation (10) can be used for calculation first and then adjusted through actual measurement.(c)RRT* generates progressively optimal subpaths within corridors and reduces path length through rewiring.Output: Smooth, near-optimal continuous paths. To verify the rationality of the selection of bias sampling probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, we conducted a detailed sensitivity analysis. While keeping other parameters unchanged, we evaluated the impact of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} ranging from 50% to 90% on algorithm performance in the same test scenario. The results are shown in Table 5. Analysis shows that when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, the algorithm achieves the best balance in planning time (135 ms), path length (25.8 m), and success rate (100%). When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is lower than 70% (such as 50%, 60%), due to insufficient sampling in the safe corridor, the algorithm requires more time for global exploration, resulting in a significant increase in planning time and a slight decline in path quality. When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is higher than 70% (such as 80%, 90%), the algorithm overly relies on the prior information of the JPS path, and its exploration ability weakens. This leads to the algorithm easily falling into local optimum when encountering obstacles that slightly deviate from the JPS path, which is manifested as a decrease in success rate (90%), an increase in the number of emergency stops (2.7 times), and a longer path length, as vehicles need to detour more circuitantly. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7 was determined as the optimal value because it maximizes the guiding role of the JPS path to achieve rapid convergence, while retaining sufficient random exploration capabilities to cope with uncertainties in the environment.Dynamic obstacle response (collaborative processing):Input: Point cloud data from external sensors.Process: Identify newly added obstacles in JPS’s global path.Initiate RRT* optimization only for affected segments (e.g. within 2 m of obstacles), keeping other paths unchanged to reduce computation.Seamlessly integrate RRT*-optimized local paths with JPS global paths for continuity.Output: Updated obstacle-avoiding path.4.Parameter configuration method for local path planning: A total of three unmanned vehicles with identical hardware were tested in this project. Due to slight mechanical and sensor tolerances, the same parameter set produced different planning behaviors across platforms. Therefore, the parameter configuration for local path planning was determined through a systematic grid search rather than empirical tuning alone. The tuning focused on parameters that most strongly influence local planner behavior, including dt_ref, max_global_plan_lookahead_dist, inflation_dist, and weight_obstacle. The objective function for optimization was defined as: 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} J = \alpha _1 T_{\text {plan}} + \alpha _2 L_{\text {dev}} + \alpha _3 I_{\text {collision}}, \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{\text {plan}}$$\end{document} is the average planning time (ms), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{\text {dev}}$$\end{document} is the maximum lateral deviation (m), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{\text {collision}}$$\end{document} is a binary indicator of collision. To prioritize planning efficiency while ensuring safety, the weights were set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _1 = 0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _2 = 0.3$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _3 = 0.2$$\end{document}. All evaluations were conducted in a simulation environment replicating the 8 m \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 6 m physical test field. The grid search for each parameter was performed while holding the remaining parameters fixed at their baseline values. The tested ranges and resulting optimal values are summarized below:dt_ref (Time resolution, 0.05–0.5 s): Smaller values (<0.1 s) significantly increased CPU load on the Raspberry Pi 4B (above 80%), while larger values (>0.3 s) resulted in insufficient control update frequency and larger lateral deviation (>0.15 m). The optimal value dt_ref=0.2 provided a good trade-off between responsiveness and computational efficiency (CPU usage \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}45%).max_global_plan_lookahead_dist (Lookahead distance, 5–12 m): Short lookahead values (<6 m) caused hesitation near obstacles and increased path replanning rate, raising planning time by over 30%. Excessively long values (>10 m) incorporated stale map information, leading to inefficient trajectory generation. A lookahead distance of 8 m was found to align well with the vehicle’s sensor range, ensuring smooth forward planning without instability.inflation_dist & weight_obstacle (Obstacle-handling parameters): These two parameters were tuned jointly due to their coupled effect on safe navigation. Small inflation_dist (<0.2 m) or low weight_obstacle (<100) increased collision risk at corners. In contrast, large inflation distances (>0.4 m) or overly high obstacle weights (>200) caused excessively conservative motion, occasional deadlocks, and unnecessary detours. The pair inflation_dist=0.25 and weight_obstacle=150 consistently yielded safe paths with minimum 0.25 m clearance while preserving planning efficiency. Overall, the final parameter set represents the Pareto-optimal solution obtained from the grid search, balancing safety, planning speed, and path smoothness across all three unmanned vehicles.5.Criteria for triggering the JPS-RRT algorithm: In a small-scale environment, the threshold Settings we make for it can be relatively small. For instance, in a pharmacy environment, when a dynamic obstacle is detected within a 2-meter radius of the vehicle’s current path segment, or when the local obstacle density measured by the LiDAR sensor exceeds the threshold of 30%, a switch from the global JPS path to the local RRT* optimization will be triggered. The selection of these threshold values (2-meter radius and 30% obstacle density) was determined through a systematic sensitivity analysis aimed at balancing planning efficiency and collision avoidance. We evaluated multiple candidate thresholds in the simulated pharmacy environment (8 m \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 6 m) under dynamic obstacle scenarios, with performance measured by three key metrics:Planning Time: Average time for local replanning (ms).Emergency Stops: Number of incidents where the vehicle stalled due to obstacle proximity (see Table 5 for definition).Path Smoothness: Average curvature change (1/m). The threshold optimization process was as follows: Radius Threshold Sweep: Obstacle detection radii from 1.0 m to 3.5 m (in increments of 0.5 m) were evaluated while fixing the obstacle density at 30%. A smaller radius (<1.5 m) caused delayed reactions, resulting in frequent emergency stops (up to 8 per run). A larger radius (>2.5 m) triggered unnecessary replanning, increasing the average planning time by more than 40%. A radius of 2.0 m provided the best trade-off, reducing emergency stops to 1–2 per run without a significant increase in computation time.Obstacle Density Threshold Sweep: Density thresholds from 20% to 50% (in steps of 5%) were tested with the radius fixed at 2.0 m. Low thresholds (<25%) triggered premature replanning in cluttered regions, increasing CPU usage by approximately 15%. High thresholds (>35%) delayed responses to dense obstacle clusters, increasing collision risk. A 30% obstacle density yielded the fewest false triggers while minimizing missed detections.Joint Validation: The threshold pair (2 m, 30%) was validated through 30 experimental runs, achieving a 95% success rate in dynamic obstacle avoidance while maintaining an average replanning time below 150 ms.These thresholds are environment-dependent; for larger operational areas they may require scaling. However, for typical pharmacy layouts, this configuration ensures robust performance without overloading the embedded computing platform.As shown in Algorithm 2, here is the pseudo-code of the fusion algorithm. Global path generation (JPS-dominant):Input: Grid map (from SLAM), start and goal points.Process: Global path generation (JPS-dominant):Input: Grid map (from SLAM), start and goal points.Process: Input: Grid map (from SLAM), start and goal points. Input: Grid map (from SLAM), start and goal points. Process: Process: JPS quickly prunes symmetric paths in the grid map, generating an initial path composed of jump points. JPS quickly prunes symmetric paths in the grid map, generating an initial path composed of jump points. The JPS path is segmented by curvature or obstacle density, with regions needing optimization marked (e.g. narrow passages, complex turns). The JPS path is segmented by curvature or obstacle density, with regions needing optimization marked (e.g. narrow passages, complex turns). Output: Line segments and optimization requirement markers. Output: Line segments and optimization requirement markers. Local optimization (RRT*-dominant):Input: JPS path segments, continuous space map, optimization markers.Process: Map JPS segments to continuous space, generating adjustable-width safety corridors:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W = W_\text {base} + \alpha \cdot \text {obstacle}_\text {density} \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}}$$\end{document} is the base width and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} is a tuning factor. The parameters in Eq. (10) were selected based on a combination of physical constraints and systematic optimization objectives. The base width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}} = 1.0\,\textrm{m}$$\end{document} is derived from the vehicle’s physical width (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {vehicle}} \approx 0.5\,\textrm{m}$$\end{document}) plus a necessary safety margin on both sides. This margin accounts for positioning errors (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 0.1\,\textrm{m}$$\end{document}, see Sect. Test results for Dijkstra and A* algorithms) and tracking control deviations, and is consistent with the common practice of setting an inflation distance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2 \times 0.25\,\textrm{m}$$\end{document}) around the robot. Formally, 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W_{\text {base}} \ge W_{\text {vehicle}} + 2 \times (\text {Positioning Error} + \text {Control Buffer}), \end{aligned}$$\end{document} ensuring a minimum clearance for safe passage. The tuning factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was determined by formulating the corridor width selection as an optimization problem with two competing objectives: (1) maximizing path feasibility (favored by a wider corridor), and (2) minimizing computational load (favored by a narrower corridor). A wider corridor increases the probability that RRT* can find a feasible path quickly but also enlarges the sampling space, potentially increasing planning time. We quantified these effects through a grid search over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \in [0.2, 1.0]$$\end{document} in our simulation environment. The primary optimization criteria were to achieve zero collisions while minimizing the planning time. The value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was found to be the Pareto-optimal point that best balanced these objectives for our specific pharmacy environment, leading to a significant reduction in planning time compared to a fixed-width corridor (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0$$\end{document}), without incurring additional collision risk. This approach provides a more principled rationale beyond pure empirical observation. For the value range of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}, we generally refer to the following range and then make specific adjustments based on the actual situation on site.Indoor/Narrow passage 0.5-1.0.Outdoor/low-density scenes: 0.2-0.5.Dynamic scenes (such as pedestrians) 0.8-1.2.(b)Sample with high probability (e.g. 70%) within safety corridors to guide RRT* tree growth along JPS paths. Sampling probability distribution: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_{\text {bias}} = 0.7 \text { (corridor)}, \quad p_{\text {uniform}} = 0.3 \text { (full space)} \end{aligned}$$\end{document} A sampling rate exceeding 70% can accelerate convergence, but it is prone to getting stuck in local optima or dead zones; a rate lower than 70% leads to scattered sampling, increased computational load, and slower convergence. Therefore, a sampling rate of 70% is selected to balance smooth path planning and response to sudden obstacles. The width of the safety corridor needs to be designed based on the actual scenario; it should not be too narrow, as it is difficult to plan, and it should not be too wide, as it increases the computational load. Equation (10) can be used for calculation first and then adjusted through actual measurement.(c)RRT* generates progressively optimal subpaths within corridors and reduces path length through rewiring.Output: Smooth, near-optimal continuous paths. To verify the rationality of the selection of bias sampling probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, we conducted a detailed sensitivity analysis. While keeping other parameters unchanged, we evaluated the impact of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} ranging from 50% to 90% on algorithm performance in the same test scenario. The results are shown in Table 5. Analysis shows that when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, the algorithm achieves the best balance in planning time (135 ms), path length (25.8 m), and success rate (100%). When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is lower than 70% (such as 50%, 60%), due to insufficient sampling in the safe corridor, the algorithm requires more time for global exploration, resulting in a significant increase in planning time and a slight decline in path quality. When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is higher than 70% (such as 80%, 90%), the algorithm overly relies on the prior information of the JPS path, and its exploration ability weakens. This leads to the algorithm easily falling into local optimum when encountering obstacles that slightly deviate from the JPS path, which is manifested as a decrease in success rate (90%), an increase in the number of emergency stops (2.7 times), and a longer path length, as vehicles need to detour more circuitantly. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7 was determined as the optimal value because it maximizes the guiding role of the JPS path to achieve rapid convergence, while retaining sufficient random exploration capabilities to cope with uncertainties in the environment. Local optimization (RRT*-dominant):Input: JPS path segments, continuous space map, optimization markers.Process: Map JPS segments to continuous space, generating adjustable-width safety corridors:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W = W_\text {base} + \alpha \cdot \text {obstacle}_\text {density} \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}}$$\end{document} is the base width and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} is a tuning factor. The parameters in Eq. (10) were selected based on a combination of physical constraints and systematic optimization objectives. The base width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}} = 1.0\,\textrm{m}$$\end{document} is derived from the vehicle’s physical width (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {vehicle}} \approx 0.5\,\textrm{m}$$\end{document}) plus a necessary safety margin on both sides. This margin accounts for positioning errors (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 0.1\,\textrm{m}$$\end{document}, see Sect. Test results for Dijkstra and A* algorithms) and tracking control deviations, and is consistent with the common practice of setting an inflation distance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2 \times 0.25\,\textrm{m}$$\end{document}) around the robot. Formally, 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W_{\text {base}} \ge W_{\text {vehicle}} + 2 \times (\text {Positioning Error} + \text {Control Buffer}), \end{aligned}$$\end{document} ensuring a minimum clearance for safe passage. The tuning factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was determined by formulating the corridor width selection as an optimization problem with two competing objectives: (1) maximizing path feasibility (favored by a wider corridor), and (2) minimizing computational load (favored by a narrower corridor). A wider corridor increases the probability that RRT* can find a feasible path quickly but also enlarges the sampling space, potentially increasing planning time. We quantified these effects through a grid search over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \in [0.2, 1.0]$$\end{document} in our simulation environment. The primary optimization criteria were to achieve zero collisions while minimizing the planning time. The value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was found to be the Pareto-optimal point that best balanced these objectives for our specific pharmacy environment, leading to a significant reduction in planning time compared to a fixed-width corridor (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0$$\end{document}), without incurring additional collision risk. This approach provides a more principled rationale beyond pure empirical observation. For the value range of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}, we generally refer to the following range and then make specific adjustments based on the actual situation on site.Indoor/Narrow passage 0.5-1.0.Outdoor/low-density scenes: 0.2-0.5.Dynamic scenes (such as pedestrians) 0.8-1.2.(b)Sample with high probability (e.g. 70%) within safety corridors to guide RRT* tree growth along JPS paths. Sampling probability distribution: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_{\text {bias}} = 0.7 \text { (corridor)}, \quad p_{\text {uniform}} = 0.3 \text { (full space)} \end{aligned}$$\end{document} A sampling rate exceeding 70% can accelerate convergence, but it is prone to getting stuck in local optima or dead zones; a rate lower than 70% leads to scattered sampling, increased computational load, and slower convergence. Therefore, a sampling rate of 70% is selected to balance smooth path planning and response to sudden obstacles. The width of the safety corridor needs to be designed based on the actual scenario; it should not be too narrow, as it is difficult to plan, and it should not be too wide, as it increases the computational load. Equation (10) can be used for calculation first and then adjusted through actual measurement.(c)RRT* generates progressively optimal subpaths within corridors and reduces path length through rewiring.Output: Smooth, near-optimal continuous paths. To verify the rationality of the selection of bias sampling probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, we conducted a detailed sensitivity analysis. While keeping other parameters unchanged, we evaluated the impact of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} ranging from 50% to 90% on algorithm performance in the same test scenario. The results are shown in Table 5. Analysis shows that when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, the algorithm achieves the best balance in planning time (135 ms), path length (25.8 m), and success rate (100%). When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is lower than 70% (such as 50%, 60%), due to insufficient sampling in the safe corridor, the algorithm requires more time for global exploration, resulting in a significant increase in planning time and a slight decline in path quality. When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is higher than 70% (such as 80%, 90%), the algorithm overly relies on the prior information of the JPS path, and its exploration ability weakens. This leads to the algorithm easily falling into local optimum when encountering obstacles that slightly deviate from the JPS path, which is manifested as a decrease in success rate (90%), an increase in the number of emergency stops (2.7 times), and a longer path length, as vehicles need to detour more circuitantly. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7 was determined as the optimal value because it maximizes the guiding role of the JPS path to achieve rapid convergence, while retaining sufficient random exploration capabilities to cope with uncertainties in the environment. Input: JPS path segments, continuous space map, optimization markers. Input: JPS path segments, continuous space map, optimization markers. Process: Map JPS segments to continuous space, generating adjustable-width safety corridors:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W = W_\text {base} + \alpha \cdot \text {obstacle}_\text {density} \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}}$$\end{document} is the base width and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} is a tuning factor. The parameters in Eq. (10) were selected based on a combination of physical constraints and systematic optimization objectives. The base width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}} = 1.0\,\textrm{m}$$\end{document} is derived from the vehicle’s physical width (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {vehicle}} \approx 0.5\,\textrm{m}$$\end{document}) plus a necessary safety margin on both sides. This margin accounts for positioning errors (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 0.1\,\textrm{m}$$\end{document}, see Sect. Test results for Dijkstra and A* algorithms) and tracking control deviations, and is consistent with the common practice of setting an inflation distance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2 \times 0.25\,\textrm{m}$$\end{document}) around the robot. Formally, 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W_{\text {base}} \ge W_{\text {vehicle}} + 2 \times (\text {Positioning Error} + \text {Control Buffer}), \end{aligned}$$\end{document} ensuring a minimum clearance for safe passage. The tuning factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was determined by formulating the corridor width selection as an optimization problem with two competing objectives: (1) maximizing path feasibility (favored by a wider corridor), and (2) minimizing computational load (favored by a narrower corridor). A wider corridor increases the probability that RRT* can find a feasible path quickly but also enlarges the sampling space, potentially increasing planning time. We quantified these effects through a grid search over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \in [0.2, 1.0]$$\end{document} in our simulation environment. The primary optimization criteria were to achieve zero collisions while minimizing the planning time. The value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was found to be the Pareto-optimal point that best balanced these objectives for our specific pharmacy environment, leading to a significant reduction in planning time compared to a fixed-width corridor (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0$$\end{document}), without incurring additional collision risk. This approach provides a more principled rationale beyond pure empirical observation. For the value range of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}, we generally refer to the following range and then make specific adjustments based on the actual situation on site.Indoor/Narrow passage 0.5-1.0.Outdoor/low-density scenes: 0.2-0.5.Dynamic scenes (such as pedestrians) 0.8-1.2.(b)Sample with high probability (e.g. 70%) within safety corridors to guide RRT* tree growth along JPS paths. Sampling probability distribution: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_{\text {bias}} = 0.7 \text { (corridor)}, \quad p_{\text {uniform}} = 0.3 \text { (full space)} \end{aligned}$$\end{document} A sampling rate exceeding 70% can accelerate convergence, but it is prone to getting stuck in local optima or dead zones; a rate lower than 70% leads to scattered sampling, increased computational load, and slower convergence. Therefore, a sampling rate of 70% is selected to balance smooth path planning and response to sudden obstacles. The width of the safety corridor needs to be designed based on the actual scenario; it should not be too narrow, as it is difficult to plan, and it should not be too wide, as it increases the computational load. Equation (10) can be used for calculation first and then adjusted through actual measurement.(c)RRT* generates progressively optimal subpaths within corridors and reduces path length through rewiring. Process: Map JPS segments to continuous space, generating adjustable-width safety corridors:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W = W_\text {base} + \alpha \cdot \text {obstacle}_\text {density} \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}}$$\end{document} is the base width and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} is a tuning factor. The parameters in Eq. (10) were selected based on a combination of physical constraints and systematic optimization objectives. The base width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {base}} = 1.0\,\textrm{m}$$\end{document} is derived from the vehicle’s physical width (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{\text {vehicle}} \approx 0.5\,\textrm{m}$$\end{document}) plus a necessary safety margin on both sides. This margin accounts for positioning errors (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 0.1\,\textrm{m}$$\end{document}, see Sect. Test results for Dijkstra and A* algorithms) and tracking control deviations, and is consistent with the common practice of setting an inflation distance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2 \times 0.25\,\textrm{m}$$\end{document}) around the robot. Formally, 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W_{\text {base}} \ge W_{\text {vehicle}} + 2 \times (\text {Positioning Error} + \text {Control Buffer}), \end{aligned}$$\end{document} ensuring a minimum clearance for safe passage. The tuning factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was determined by formulating the corridor width selection as an optimization problem with two competing objectives: (1) maximizing path feasibility (favored by a wider corridor), and (2) minimizing computational load (favored by a narrower corridor). A wider corridor increases the probability that RRT* can find a feasible path quickly but also enlarges the sampling space, potentially increasing planning time. We quantified these effects through a grid search over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \in [0.2, 1.0]$$\end{document} in our simulation environment. The primary optimization criteria were to achieve zero collisions while minimizing the planning time. The value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.5$$\end{document} was found to be the Pareto-optimal point that best balanced these objectives for our specific pharmacy environment, leading to a significant reduction in planning time compared to a fixed-width corridor (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0$$\end{document}), without incurring additional collision risk. This approach provides a more principled rationale beyond pure empirical observation. For the value range of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}, we generally refer to the following range and then make specific adjustments based on the actual situation on site.Indoor/Narrow passage 0.5-1.0.Outdoor/low-density scenes: 0.2-0.5.Dynamic scenes (such as pedestrians) 0.8-1.2.(b)Sample with high probability (e.g. 70%) within safety corridors to guide RRT* tree growth along JPS paths. Sampling probability distribution: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_{\text {bias}} = 0.7 \text { (corridor)}, \quad p_{\text {uniform}} = 0.3 \text { (full space)} \end{aligned}$$\end{document} A sampling rate exceeding 70% can accelerate convergence, but it is prone to getting stuck in local optima or dead zones; a rate lower than 70% leads to scattered sampling, increased computational load, and slower convergence. Therefore, a sampling rate of 70% is selected to balance smooth path planning and response to sudden obstacles. The width of the safety corridor needs to be designed based on the actual scenario; it should not be too narrow, as it is difficult to plan, and it should not be too wide, as it increases the computational load. Equation (10) can be used for calculation first and then adjusted through actual measurement.(c)RRT* generates progressively optimal subpaths within corridors and reduces path length through rewiring. Map JPS segments to continuous space, generating adjustable-width safety corridors: Map JPS segments to continuous space, generating adjustable-width safety corridors: Indoor/Narrow passage 0.5-1.0. Indoor/Narrow passage 0.5-1.0. Outdoor/low-density scenes: 0.2-0.5. Outdoor/low-density scenes: 0.2-0.5. Dynamic scenes (such as pedestrians) 0.8-1.2. Dynamic scenes (such as pedestrians) 0.8-1.2. (b)Sample with high probability (e.g. 70%) within safety corridors to guide RRT* tree growth along JPS paths. Sampling probability distribution: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_{\text {bias}} = 0.7 \text { (corridor)}, \quad p_{\text {uniform}} = 0.3 \text { (full space)} \end{aligned}$$\end{document} A sampling rate exceeding 70% can accelerate convergence, but it is prone to getting stuck in local optima or dead zones; a rate lower than 70% leads to scattered sampling, increased computational load, and slower convergence. Therefore, a sampling rate of 70% is selected to balance smooth path planning and response to sudden obstacles. The width of the safety corridor needs to be designed based on the actual scenario; it should not be too narrow, as it is difficult to plan, and it should not be too wide, as it increases the computational load. Equation (10) can be used for calculation first and then adjusted through actual measurement. Sample with high probability (e.g. 70%) within safety corridors to guide RRT* tree growth along JPS paths. Sampling probability distribution: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_{\text {bias}} = 0.7 \text { (corridor)}, \quad p_{\text {uniform}} = 0.3 \text { (full space)} \end{aligned}$$\end{document} A sampling rate exceeding 70% can accelerate convergence, but it is prone to getting stuck in local optima or dead zones; a rate lower than 70% leads to scattered sampling, increased computational load, and slower convergence. Therefore, a sampling rate of 70% is selected to balance smooth path planning and response to sudden obstacles. The width of the safety corridor needs to be designed based on the actual scenario; it should not be too narrow, as it is difficult to plan, and it should not be too wide, as it increases the computational load. Equation (10) can be used for calculation first and then adjusted through actual measurement. (c)RRT* generates progressively optimal subpaths within corridors and reduces path length through rewiring. RRT* generates progressively optimal subpaths within corridors and reduces path length through rewiring. Output: Smooth, near-optimal continuous paths. To verify the rationality of the selection of bias sampling probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, we conducted a detailed sensitivity analysis. While keeping other parameters unchanged, we evaluated the impact of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} ranging from 50% to 90% on algorithm performance in the same test scenario. The results are shown in Table 5. Analysis shows that when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, the algorithm achieves the best balance in planning time (135 ms), path length (25.8 m), and success rate (100%). When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is lower than 70% (such as 50%, 60%), due to insufficient sampling in the safe corridor, the algorithm requires more time for global exploration, resulting in a significant increase in planning time and a slight decline in path quality. When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is higher than 70% (such as 80%, 90%), the algorithm overly relies on the prior information of the JPS path, and its exploration ability weakens. This leads to the algorithm easily falling into local optimum when encountering obstacles that slightly deviate from the JPS path, which is manifested as a decrease in success rate (90%), an increase in the number of emergency stops (2.7 times), and a longer path length, as vehicles need to detour more circuitantly. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7 was determined as the optimal value because it maximizes the guiding role of the JPS path to achieve rapid convergence, while retaining sufficient random exploration capabilities to cope with uncertainties in the environment. Output: Smooth, near-optimal continuous paths. To verify the rationality of the selection of bias sampling probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, we conducted a detailed sensitivity analysis. While keeping other parameters unchanged, we evaluated the impact of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} ranging from 50% to 90% on algorithm performance in the same test scenario. The results are shown in Table 5. Analysis shows that when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7, the algorithm achieves the best balance in planning time (135 ms), path length (25.8 m), and success rate (100%). When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is lower than 70% (such as 50%, 60%), due to insufficient sampling in the safe corridor, the algorithm requires more time for global exploration, resulting in a significant increase in planning time and a slight decline in path quality. When the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} is higher than 70% (such as 80%, 90%), the algorithm overly relies on the prior information of the JPS path, and its exploration ability weakens. This leads to the algorithm easily falling into local optimum when encountering obstacles that slightly deviate from the JPS path, which is manifested as a decrease in success rate (90%), an increase in the number of emergency stops (2.7 times), and a longer path length, as vehicles need to detour more circuitantly. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document}=0.7 was determined as the optimal value because it maximizes the guiding role of the JPS path to achieve rapid convergence, while retaining sufficient random exploration capabilities to cope with uncertainties in the environment. Dynamic obstacle response (collaborative processing):Input: Point cloud data from external sensors.Process: Dynamic obstacle response (collaborative processing):Input: Point cloud data from external sensors.Process: Input: Point cloud data from external sensors. Input: Point cloud data from external sensors. Process: Process: Identify newly added obstacles in JPS’s global path. Identify newly added obstacles in JPS’s global path. Initiate RRT* optimization only for affected segments (e.g. within 2 m of obstacles), keeping other paths unchanged to reduce computation. Initiate RRT* optimization only for affected segments (e.g. within 2 m of obstacles), keeping other paths unchanged to reduce computation. Seamlessly integrate RRT*-optimized local paths with JPS global paths for continuity. Seamlessly integrate RRT*-optimized local paths with JPS global paths for continuity. Output: Updated obstacle-avoiding path. Output: Updated obstacle-avoiding path. 4.Parameter configuration method for local path planning: A total of three unmanned vehicles with identical hardware were tested in this project. Due to slight mechanical and sensor tolerances, the same parameter set produced different planning behaviors across platforms. Therefore, the parameter configuration for local path planning was determined through a systematic grid search rather than empirical tuning alone. The tuning focused on parameters that most strongly influence local planner behavior, including dt_ref, max_global_plan_lookahead_dist, inflation_dist, and weight_obstacle. The objective function for optimization was defined as: 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} J = \alpha _1 T_{\text {plan}} + \alpha _2 L_{\text {dev}} + \alpha _3 I_{\text {collision}}, \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{\text {plan}}$$\end{document} is the average planning time (ms), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{\text {dev}}$$\end{document} is the maximum lateral deviation (m), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{\text {collision}}$$\end{document} is a binary indicator of collision. To prioritize planning efficiency while ensuring safety, the weights were set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _1 = 0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _2 = 0.3$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _3 = 0.2$$\end{document}. All evaluations were conducted in a simulation environment replicating the 8 m \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 6 m physical test field. The grid search for each parameter was performed while holding the remaining parameters fixed at their baseline values. The tested ranges and resulting optimal values are summarized below:dt_ref (Time resolution, 0.05–0.5 s): Smaller values (<0.1 s) significantly increased CPU load on the Raspberry Pi 4B (above 80%), while larger values (>0.3 s) resulted in insufficient control update frequency and larger lateral deviation (>0.15 m). The optimal value dt_ref=0.2 provided a good trade-off between responsiveness and computational efficiency (CPU usage \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}45%).max_global_plan_lookahead_dist (Lookahead distance, 5–12 m): Short lookahead values (<6 m) caused hesitation near obstacles and increased path replanning rate, raising planning time by over 30%. Excessively long values (>10 m) incorporated stale map information, leading to inefficient trajectory generation. A lookahead distance of 8 m was found to align well with the vehicle’s sensor range, ensuring smooth forward planning without instability.inflation_dist & weight_obstacle (Obstacle-handling parameters): These two parameters were tuned jointly due to their coupled effect on safe navigation. Small inflation_dist (<0.2 m) or low weight_obstacle (<100) increased collision risk at corners. In contrast, large inflation distances (>0.4 m) or overly high obstacle weights (>200) caused excessively conservative motion, occasional deadlocks, and unnecessary detours. The pair inflation_dist=0.25 and weight_obstacle=150 consistently yielded safe paths with minimum 0.25 m clearance while preserving planning efficiency. Overall, the final parameter set represents the Pareto-optimal solution obtained from the grid search, balancing safety, planning speed, and path smoothness across all three unmanned vehicles. Parameter configuration method for local path planning: A total of three unmanned vehicles with identical hardware were tested in this project. Due to slight mechanical and sensor tolerances, the same parameter set produced different planning behaviors across platforms. Therefore, the parameter configuration for local path planning was determined through a systematic grid search rather than empirical tuning alone. The tuning focused on parameters that most strongly influence local planner behavior, including dt_ref, max_global_plan_lookahead_dist, inflation_dist, and weight_obstacle. The objective function for optimization was defined as: 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} J = \alpha _1 T_{\text {plan}} + \alpha _2 L_{\text {dev}} + \alpha _3 I_{\text {collision}}, \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{\text {plan}}$$\end{document} is the average planning time (ms), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{\text {dev}}$$\end{document} is the maximum lateral deviation (m), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{\text {collision}}$$\end{document} is a binary indicator of collision. To prioritize planning efficiency while ensuring safety, the weights were set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _1 = 0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _2 = 0.3$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _3 = 0.2$$\end{document}. All evaluations were conducted in a simulation environment replicating the 8 m \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 6 m physical test field. The grid search for each parameter was performed while holding the remaining parameters fixed at their baseline values. The tested ranges and resulting optimal values are summarized below:dt_ref (Time resolution, 0.05–0.5 s): Smaller values (<0.1 s) significantly increased CPU load on the Raspberry Pi 4B (above 80%), while larger values (>0.3 s) resulted in insufficient control update frequency and larger lateral deviation (>0.15 m). The optimal value dt_ref=0.2 provided a good trade-off between responsiveness and computational efficiency (CPU usage \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}45%).max_global_plan_lookahead_dist (Lookahead distance, 5–12 m): Short lookahead values (<6 m) caused hesitation near obstacles and increased path replanning rate, raising planning time by over 30%. Excessively long values (>10 m) incorporated stale map information, leading to inefficient trajectory generation. A lookahead distance of 8 m was found to align well with the vehicle’s sensor range, ensuring smooth forward planning without instability.inflation_dist & weight_obstacle (Obstacle-handling parameters): These two parameters were tuned jointly due to their coupled effect on safe navigation. Small inflation_dist (<0.2 m) or low weight_obstacle (<100) increased collision risk at corners. In contrast, large inflation distances (>0.4 m) or overly high obstacle weights (>200) caused excessively conservative motion, occasional deadlocks, and unnecessary detours. The pair inflation_dist=0.25 and weight_obstacle=150 consistently yielded safe paths with minimum 0.25 m clearance while preserving planning efficiency. Overall, the final parameter set represents the Pareto-optimal solution obtained from the grid search, balancing safety, planning speed, and path smoothness across all three unmanned vehicles. dt_ref (Time resolution, 0.05–0.5 s): Smaller values (<0.1 s) significantly increased CPU load on the Raspberry Pi 4B (above 80%), while larger values (>0.3 s) resulted in insufficient control update frequency and larger lateral deviation (>0.15 m). The optimal value dt_ref=0.2 provided a good trade-off between responsiveness and computational efficiency (CPU usage \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}45%). dt_ref (Time resolution, 0.05–0.5 s): Smaller values (<0.1 s) significantly increased CPU load on the Raspberry Pi 4B (above 80%), while larger values (>0.3 s) resulted in insufficient control update frequency and larger lateral deviation (>0.15 m). The optimal value dt_ref=0.2 provided a good trade-off between responsiveness and computational efficiency (CPU usage \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}45%). max_global_plan_lookahead_dist (Lookahead distance, 5–12 m): Short lookahead values (<6 m) caused hesitation near obstacles and increased path replanning rate, raising planning time by over 30%. Excessively long values (>10 m) incorporated stale map information, leading to inefficient trajectory generation. A lookahead distance of 8 m was found to align well with the vehicle’s sensor range, ensuring smooth forward planning without instability. max_global_plan_lookahead_dist (Lookahead distance, 5–12 m): Short lookahead values (<6 m) caused hesitation near obstacles and increased path replanning rate, raising planning time by over 30%. Excessively long values (>10 m) incorporated stale map information, leading to inefficient trajectory generation. A lookahead distance of 8 m was found to align well with the vehicle’s sensor range, ensuring smooth forward planning without instability. inflation_dist & weight_obstacle (Obstacle-handling parameters): These two parameters were tuned jointly due to their coupled effect on safe navigation. Small inflation_dist (<0.2 m) or low weight_obstacle (<100) increased collision risk at corners. In contrast, large inflation distances (>0.4 m) or overly high obstacle weights (>200) caused excessively conservative motion, occasional deadlocks, and unnecessary detours. The pair inflation_dist=0.25 and weight_obstacle=150 consistently yielded safe paths with minimum 0.25 m clearance while preserving planning efficiency. inflation_dist & weight_obstacle (Obstacle-handling parameters): These two parameters were tuned jointly due to their coupled effect on safe navigation. Small inflation_dist (<0.2 m) or low weight_obstacle (<100) increased collision risk at corners. In contrast, large inflation distances (>0.4 m) or overly high obstacle weights (>200) caused excessively conservative motion, occasional deadlocks, and unnecessary detours. The pair inflation_dist=0.25 and weight_obstacle=150 consistently yielded safe paths with minimum 0.25 m clearance while preserving planning efficiency. 5.Criteria for triggering the JPS-RRT algorithm: In a small-scale environment, the threshold Settings we make for it can be relatively small. For instance, in a pharmacy environment, when a dynamic obstacle is detected within a 2-meter radius of the vehicle’s current path segment, or when the local obstacle density measured by the LiDAR sensor exceeds the threshold of 30%, a switch from the global JPS path to the local RRT* optimization will be triggered. The selection of these threshold values (2-meter radius and 30% obstacle density) was determined through a systematic sensitivity analysis aimed at balancing planning efficiency and collision avoidance. We evaluated multiple candidate thresholds in the simulated pharmacy environment (8 m \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 6 m) under dynamic obstacle scenarios, with performance measured by three key metrics:Planning Time: Average time for local replanning (ms).Emergency Stops: Number of incidents where the vehicle stalled due to obstacle proximity (see Table 5 for definition).Path Smoothness: Average curvature change (1/m). The threshold optimization process was as follows: Radius Threshold Sweep: Obstacle detection radii from 1.0 m to 3.5 m (in increments of 0.5 m) were evaluated while fixing the obstacle density at 30%. A smaller radius (<1.5 m) caused delayed reactions, resulting in frequent emergency stops (up to 8 per run). A larger radius (>2.5 m) triggered unnecessary replanning, increasing the average planning time by more than 40%. A radius of 2.0 m provided the best trade-off, reducing emergency stops to 1–2 per run without a significant increase in computation time.Obstacle Density Threshold Sweep: Density thresholds from 20% to 50% (in steps of 5%) were tested with the radius fixed at 2.0 m. Low thresholds (<25%) triggered premature replanning in cluttered regions, increasing CPU usage by approximately 15%. High thresholds (>35%) delayed responses to dense obstacle clusters, increasing collision risk. A 30% obstacle density yielded the fewest false triggers while minimizing missed detections.Joint Validation: The threshold pair (2 m, 30%) was validated through 30 experimental runs, achieving a 95% success rate in dynamic obstacle avoidance while maintaining an average replanning time below 150 ms. Criteria for triggering the JPS-RRT algorithm: In a small-scale environment, the threshold Settings we make for it can be relatively small. For instance, in a pharmacy environment, when a dynamic obstacle is detected within a 2-meter radius of the vehicle’s current path segment, or when the local obstacle density measured by the LiDAR sensor exceeds the threshold of 30%, a switch from the global JPS path to the local RRT* optimization will be triggered. The selection of these threshold values (2-meter radius and 30% obstacle density) was determined through a systematic sensitivity analysis aimed at balancing planning efficiency and collision avoidance. We evaluated multiple candidate thresholds in the simulated pharmacy environment (8 m \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 6 m) under dynamic obstacle scenarios, with performance measured by three key metrics:Planning Time: Average time for local replanning (ms).Emergency Stops: Number of incidents where the vehicle stalled due to obstacle proximity (see Table 5 for definition).Path Smoothness: Average curvature change (1/m). The threshold optimization process was as follows: Radius Threshold Sweep: Obstacle detection radii from 1.0 m to 3.5 m (in increments of 0.5 m) were evaluated while fixing the obstacle density at 30%. A smaller radius (<1.5 m) caused delayed reactions, resulting in frequent emergency stops (up to 8 per run). A larger radius (>2.5 m) triggered unnecessary replanning, increasing the average planning time by more than 40%. A radius of 2.0 m provided the best trade-off, reducing emergency stops to 1–2 per run without a significant increase in computation time.Obstacle Density Threshold Sweep: Density thresholds from 20% to 50% (in steps of 5%) were tested with the radius fixed at 2.0 m. Low thresholds (<25%) triggered premature replanning in cluttered regions, increasing CPU usage by approximately 15%. High thresholds (>35%) delayed responses to dense obstacle clusters, increasing collision risk. A 30% obstacle density yielded the fewest false triggers while minimizing missed detections.Joint Validation: The threshold pair (2 m, 30%) was validated through 30 experimental runs, achieving a 95% success rate in dynamic obstacle avoidance while maintaining an average replanning time below 150 ms. Planning Time: Average time for local replanning (ms). Planning Time: Average time for local replanning (ms). Emergency Stops: Number of incidents where the vehicle stalled due to obstacle proximity (see Table 5 for definition). Emergency Stops: Number of incidents where the vehicle stalled due to obstacle proximity (see Table 5 for definition). Path Smoothness: Average curvature change (1/m). Path Smoothness: Average curvature change (1/m). Radius Threshold Sweep: Obstacle detection radii from 1.0 m to 3.5 m (in increments of 0.5 m) were evaluated while fixing the obstacle density at 30%. A smaller radius (<1.5 m) caused delayed reactions, resulting in frequent emergency stops (up to 8 per run). A larger radius (>2.5 m) triggered unnecessary replanning, increasing the average planning time by more than 40%. A radius of 2.0 m provided the best trade-off, reducing emergency stops to 1–2 per run without a significant increase in computation time. Radius Threshold Sweep: Obstacle detection radii from 1.0 m to 3.5 m (in increments of 0.5 m) were evaluated while fixing the obstacle density at 30%. A smaller radius (<1.5 m) caused delayed reactions, resulting in frequent emergency stops (up to 8 per run). A larger radius (>2.5 m) triggered unnecessary replanning, increasing the average planning time by more than 40%. A radius of 2.0 m provided the best trade-off, reducing emergency stops to 1–2 per run without a significant increase in computation time. Obstacle Density Threshold Sweep: Density thresholds from 20% to 50% (in steps of 5%) were tested with the radius fixed at 2.0 m. Low thresholds (<25%) triggered premature replanning in cluttered regions, increasing CPU usage by approximately 15%. High thresholds (>35%) delayed responses to dense obstacle clusters, increasing collision risk. A 30% obstacle density yielded the fewest false triggers while minimizing missed detections. Obstacle Density Threshold Sweep: Density thresholds from 20% to 50% (in steps of 5%) were tested with the radius fixed at 2.0 m. Low thresholds (<25%) triggered premature replanning in cluttered regions, increasing CPU usage by approximately 15%. High thresholds (>35%) delayed responses to dense obstacle clusters, increasing collision risk. A 30% obstacle density yielded the fewest false triggers while minimizing missed detections. Joint Validation: The threshold pair (2 m, 30%) was validated through 30 experimental runs, achieving a 95% success rate in dynamic obstacle avoidance while maintaining an average replanning time below 150 ms. Joint Validation: The threshold pair (2 m, 30%) was validated through 30 experimental runs, achieving a 95% success rate in dynamic obstacle avoidance while maintaining an average replanning time below 150 ms. Algorithm 2JPS–RRT* hybrid path planner. Fig. 9Interaction flowchart between JPS and RRT*. Table 5Effect of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} on planning performance. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} Planning time (ms)Path length (m)Success rate (%)Emergency stops50% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$158 \pm 11$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$26.1 \pm 0.3$$\end{document} 93.3 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.3 \pm 0.9$$\end{document} 60% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$142 \pm 9$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$25.9 \pm 0.2$$\end{document} 96.7 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7 \pm 0.8$$\end{document} 70% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {135 \pm 8}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {25.8 \pm 0.2}$$\end{document} 100 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {1.0 \pm 0.6}$$\end{document} 80% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$140 \pm 10$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$26.0 \pm 0.3$$\end{document} 96.7 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.3 \pm 0.7$$\end{document} 90% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$165 \pm 12$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$26.3 \pm 0.4$$\end{document} 90.0 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.7 \pm 1.0$$\end{document} Significant values are in bold. JPS–RRT* hybrid path planner. Interaction flowchart between JPS and RRT*. Table 5Effect of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} on planning performance. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} Planning time (ms)Path length (m)Success rate (%)Emergency stops50% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$158 \pm 11$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$26.1 \pm 0.3$$\end{document} 93.3 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.3 \pm 0.9$$\end{document} 60% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$142 \pm 9$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$25.9 \pm 0.2$$\end{document} 96.7 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7 \pm 0.8$$\end{document} 70% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {135 \pm 8}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {25.8 \pm 0.2}$$\end{document} 100 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {1.0 \pm 0.6}$$\end{document} 80% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$140 \pm 10$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$26.0 \pm 0.3$$\end{document} 96.7 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.3 \pm 0.7$$\end{document} 90% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$165 \pm 12$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$26.3 \pm 0.4$$\end{document} 90.0 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.7 \pm 1.0$$\end{document} Significant values are in bold. Effect of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} on planning performance. Significant values are in bold. Detailed description of the interaction process The proposed planner integrates the global search efficiency of Jump Point Search (JPS) with the continuous optimization capability of RRT*. The interaction procedure is summarized as follows.As shown in Fig. 9, here is the flowchart of the fusion algorithm. Environment assessment : The grid map is initialized with start and goal states. The system determines whether the environment exhibits structured-road characteristics. Structured environments activate the JPS-based global search; otherwise, the planner directly performs exploratory RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}.Global planning using JPS : In structured environments, JPS generates an initial sequence of jump points. The resulting path is checked for existence and local obstacle-free feasibility. If fully feasible, the JPS path is output as the global solution.Transition to RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} : If the JPS path contains blocked segments, the planner switches to RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} for local replanning. The jump points produced by JPS are used to bias RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} sampling, increasing sampling probability along the JPS-inferred global corridor. RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} generates candidate local trajectories under this guided-sampling strategy.Fusion path evaluation : Candidate RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} paths are evaluated based on real-time feasibility, obstacle clearance, and smoothness. If a path satisfies these criteria, it is adopted as the final fusion path. If not, the sampling resolution of RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} is reduced to accelerate replanning until a feasible solution emerges.Exploratory planning in unstructured environments : When the environment is unstructured, RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} operates with uniform sampling. After each iteration, the planner checks trajectory convergence. If no significant cost improvement occurs within a predefined number of iterations, the current RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} solution is accepted; otherwise, the sampling density is increased to enhance exploration.Path output and tracking : Depending on the branch, the planner outputs either the feasible JPS path, the RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}-refined fusion path, or the optimized RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} trajectory. The selected path is then provided to the trajectory-tracking controller. This interaction strategy ensures both fast global routing via JPS and high-quality, dynamically feasible trajectories via RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}, while maintaining full algorithmic reproducibility. Environment assessment : The grid map is initialized with start and goal states. The system determines whether the environment exhibits structured-road characteristics. Structured environments activate the JPS-based global search; otherwise, the planner directly performs exploratory RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}. Environment assessment : The grid map is initialized with start and goal states. The system determines whether the environment exhibits structured-road characteristics. Structured environments activate the JPS-based global search; otherwise, the planner directly performs exploratory RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}. Global planning using JPS : In structured environments, JPS generates an initial sequence of jump points. The resulting path is checked for existence and local obstacle-free feasibility. If fully feasible, the JPS path is output as the global solution. Global planning using JPS : In structured environments, JPS generates an initial sequence of jump points. The resulting path is checked for existence and local obstacle-free feasibility. If fully feasible, the JPS path is output as the global solution. Transition to RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} : If the JPS path contains blocked segments, the planner switches to RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} for local replanning. The jump points produced by JPS are used to bias RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} sampling, increasing sampling probability along the JPS-inferred global corridor. RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} generates candidate local trajectories under this guided-sampling strategy. Transition to RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} : If the JPS path contains blocked segments, the planner switches to RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} for local replanning. The jump points produced by JPS are used to bias RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} sampling, increasing sampling probability along the JPS-inferred global corridor. RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} generates candidate local trajectories under this guided-sampling strategy. Fusion path evaluation : Candidate RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} paths are evaluated based on real-time feasibility, obstacle clearance, and smoothness. If a path satisfies these criteria, it is adopted as the final fusion path. If not, the sampling resolution of RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} is reduced to accelerate replanning until a feasible solution emerges. Fusion path evaluation : Candidate RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} paths are evaluated based on real-time feasibility, obstacle clearance, and smoothness. If a path satisfies these criteria, it is adopted as the final fusion path. If not, the sampling resolution of RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} is reduced to accelerate replanning until a feasible solution emerges. Exploratory planning in unstructured environments : When the environment is unstructured, RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} operates with uniform sampling. After each iteration, the planner checks trajectory convergence. If no significant cost improvement occurs within a predefined number of iterations, the current RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} solution is accepted; otherwise, the sampling density is increased to enhance exploration. Exploratory planning in unstructured environments : When the environment is unstructured, RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} operates with uniform sampling. After each iteration, the planner checks trajectory convergence. If no significant cost improvement occurs within a predefined number of iterations, the current RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} solution is accepted; otherwise, the sampling density is increased to enhance exploration. Path output and tracking : Depending on the branch, the planner outputs either the feasible JPS path, the RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}-refined fusion path, or the optimized RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} trajectory. The selected path is then provided to the trajectory-tracking controller. This interaction strategy ensures both fast global routing via JPS and high-quality, dynamically feasible trajectories via RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}, while maintaining full algorithmic reproducibility. Path output and tracking : Depending on the branch, the planner outputs either the feasible JPS path, the RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}-refined fusion path, or the optimized RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} trajectory. The selected path is then provided to the trajectory-tracking controller. This interaction strategy ensures both fast global routing via JPS and high-quality, dynamically feasible trajectories via RRT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document}, while maintaining full algorithmic reproducibility. Complementary process Efficient structured environment search (JPS compensates RRT*): RRT*’s random sampling causes redundant computations in regular grids (e.g. ineffective exploration in long straight corridors). JPS quickly generates static straight paths, allowing RRT* to expand nodes and verify path feasibility through local sampling without global sampling waste. Testing showed planning time reduction from 320ms (RRT*) to 80ms (JPS-RRT*) in 50\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}50 grid maps.Unstructured environment path optimization (RRT* compensates JPS): JPS paths may be infeasible in complex terrains (e.g. impassable narrow turns). Using previously marked optimization nodes, RRT* performs local optimization in complex JPS regions, generating smooth paths meeting kinematic constraints. This reduces path curvature by 60%, accommodating vehicle minimum turning radius.Efficient dynamic obstacle handling (collaborative mechanism): While JPS cannot handle dynamic obstacles and RRT* global replanning is time-consuming, their collaboration uses JPS for global reference paths with RRT* adjusting only affected local regions. This reduces dynamic obstacle response time from 500ms (pure RRT*) to 150ms. Fig. 10Path diagram of JPS-RRT* algorithm. Efficient structured environment search (JPS compensates RRT*): RRT*’s random sampling causes redundant computations in regular grids (e.g. ineffective exploration in long straight corridors). JPS quickly generates static straight paths, allowing RRT* to expand nodes and verify path feasibility through local sampling without global sampling waste. Testing showed planning time reduction from 320ms (RRT*) to 80ms (JPS-RRT*) in 50\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}50 grid maps. Efficient structured environment search (JPS compensates RRT*): RRT*’s random sampling causes redundant computations in regular grids (e.g. ineffective exploration in long straight corridors). JPS quickly generates static straight paths, allowing RRT* to expand nodes and verify path feasibility through local sampling without global sampling waste. Testing showed planning time reduction from 320ms (RRT*) to 80ms (JPS-RRT*) in 50\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}50 grid maps. Unstructured environment path optimization (RRT* compensates JPS): JPS paths may be infeasible in complex terrains (e.g. impassable narrow turns). Using previously marked optimization nodes, RRT* performs local optimization in complex JPS regions, generating smooth paths meeting kinematic constraints. This reduces path curvature by 60%, accommodating vehicle minimum turning radius. Unstructured environment path optimization (RRT* compensates JPS): JPS paths may be infeasible in complex terrains (e.g. impassable narrow turns). Using previously marked optimization nodes, RRT* performs local optimization in complex JPS regions, generating smooth paths meeting kinematic constraints. This reduces path curvature by 60%, accommodating vehicle minimum turning radius. Efficient dynamic obstacle handling (collaborative mechanism): While JPS cannot handle dynamic obstacles and RRT* global replanning is time-consuming, their collaboration uses JPS for global reference paths with RRT* adjusting only affected local regions. This reduces dynamic obstacle response time from 500ms (pure RRT*) to 150ms. Efficient dynamic obstacle handling (collaborative mechanism): While JPS cannot handle dynamic obstacles and RRT* global replanning is time-consuming, their collaboration uses JPS for global reference paths with RRT* adjusting only affected local regions. This reduces dynamic obstacle response time from 500ms (pure RRT*) to 150ms. Path diagram of JPS-RRT* algorithm. As shown in Fig. 10, here is the simulation diagram of the algorithm and relevant comparisons have been made with other algorithms. Performance comparison Testing in real scenarios yielded comparative performance data. We compared JPS-RRT* with its individual components and different planning algorithms under identical tasks. The metric Emergency Stops in Table 6 is rigorously defined by the following unambiguous, system-measured criteria. An emergency stop event is recorded if the vehicle’s state simultaneously satisfies the following three conditions for a continuous duration exceeding 2.0 seconds: Zero velocity: The linear velocity of the vehicle, as reported by the wheel encoders, drops below 0.05 m/s.Path planning failure: The path planning module (JPS, RRT, or JPS-RRT) fails to output a feasible collision-free path to the immediate sub-goal.Proximity to obstacle: The minimum distance to the nearest obstacle, measured by the LiDAR point cloud, is less than 0.3 meters.This objective definition captures events where the vehicle is effectively “trapped”, it has stopped because no safe path can be found in its immediate vicinity. The choice of a 2-second duration prevents brief pauses (e.g. for calculation or normal braking) from being misclassified as emergencies. The counts in Table 6 are the mean values automatically logged by the system software across 30 independent runs for each algorithm.Table 6Performance comparison between JPS-RRT* and its components.MetricJPSRRT*JPS-RRT*Planning Time (s)0.12.30.4Path Length (m)28.525.325.8Emergency Stops621 Zero velocity: The linear velocity of the vehicle, as reported by the wheel encoders, drops below 0.05 m/s. Zero velocity: The linear velocity of the vehicle, as reported by the wheel encoders, drops below 0.05 m/s. Path planning failure: The path planning module (JPS, RRT, or JPS-RRT) fails to output a feasible collision-free path to the immediate sub-goal. Path planning failure: The path planning module (JPS, RRT, or JPS-RRT) fails to output a feasible collision-free path to the immediate sub-goal. Proximity to obstacle: The minimum distance to the nearest obstacle, measured by the LiDAR point cloud, is less than 0.3 meters. Proximity to obstacle: The minimum distance to the nearest obstacle, measured by the LiDAR point cloud, is less than 0.3 meters. Table 6Performance comparison between JPS-RRT* and its components.MetricJPSRRT*JPS-RRT*Planning Time (s)0.12.30.4Path Length (m)28.525.325.8Emergency Stops621 Performance comparison between JPS-RRT* and its components. As shown in Table 6, the pure JPS algorithm resulted in the most emergency stops (6 times) because it cannot handle dynamic obstacles, leading to frequent deadlocks. The pure RRT* algorithm performed better (2 stops) due to its inherent ability to react to changes. The hybrid JPS-RRT* algorithm achieved the best result (1 stop) by leveraging the global efficiency of JPS and the local adaptability of RRT*, effectively minimizing unexpected stops. The main reason is that when an unmanned vehicle gets extremely close to a dynamic obstacle in a short period of time, the planned path will overlap with the obstacle, causing it to be unable to get out of the predicament in time and fall into a dead zone, which leads to an emergency stop. Different algorithms handle this situation differently, so there is a significant difference in the number of emergency stops. As shown in Table 7, the performance comparison among different algorithms is presented here.Table 7Performance comparison between JPS-RRT* and other algorithms.AlgorithmPlanning time (ms)Path length (m)Smoothness (1/m)Dijkstra62028.53.1A*8528.72.8RRT*32025.31.2JPS-RRT*13525.91.4Table 8Statistical comparison of planning algorithms.Performance indicatorsGroupMean differenceT-statisticp-valuePlan time (ms)JPS-RRT* vs. Dijkstra\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-485.0$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-512.3, -457.7]$$\end{document}35.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.8 \times 10^{-30}$$\end{document}JPS-RRT* vs. A*\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-50.0$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-56.8, -43.2]$$\end{document}15.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.1 \times 10^{-16}$$\end{document}JPS-RRT* vs. RRT*\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-185.0$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-197.4, -172.6]$$\end{document}28.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.2 \times 10^{-25}$$\end{document}Path length (m)JPS-RRT* vs. Dijkstra\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-2.6$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-2.8, -2.4]$$\end{document}22.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.1 \times 10^{-21}$$\end{document}JPS-RRT* vs. A*\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-2.8$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-3.1, -2.5]$$\end{document}18.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.3 \times 10^{-18}$$\end{document} Table 7Performance comparison between JPS-RRT* and other algorithms.AlgorithmPlanning time (ms)Path length (m)Smoothness (1/m)Dijkstra62028.53.1A*8528.72.8RRT*32025.31.2JPS-RRT*13525.91.4 Performance comparison between JPS-RRT* and other algorithms. Table 8Statistical comparison of planning algorithms.Performance indicatorsGroupMean differenceT-statisticp-valuePlan time (ms)JPS-RRT* vs. Dijkstra\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-485.0$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-512.3, -457.7]$$\end{document}35.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.8 \times 10^{-30}$$\end{document}JPS-RRT* vs. A*\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-50.0$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-56.8, -43.2]$$\end{document}15.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.1 \times 10^{-16}$$\end{document}JPS-RRT* vs. RRT*\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-185.0$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-197.4, -172.6]$$\end{document}28.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.2 \times 10^{-25}$$\end{document}Path length (m)JPS-RRT* vs. Dijkstra\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-2.6$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-2.8, -2.4]$$\end{document}22.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.1 \times 10^{-21}$$\end{document}JPS-RRT* vs. A*\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-2.8$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-3.1, -2.5]$$\end{document}18.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.3 \times 10^{-18}$$\end{document} Statistical comparison of planning algorithms. To rigorously validate the performance superiority of the proposed JPS-RRT* algorithm, a statistical analysis was conducted based on 30 independent runs. The analysis results are shown in Table 8.The results, presented as mean ± standard deviation in Table 5 and Table 6, demonstrate the stability of our method. Paired-sample t-tests confirmed that the improvements of JPS-RRT* over baseline algorithms (Dijkstra, A, and RRT) are statistically significant (p < 0.01 for all comparisons). For example, the reduction in planning time compared to RRT* yielded a t-statistic of 28.5 with a p-value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.2 \times 10^{-25}$$\end{document}. Drug image recognition Before pharmacy robots transport medications, accurate drug identification is essential for correct grasping. The overall workflow of the drug recognition process is shown in Fig. 11, outlining the stages from image acquisition and preprocessing to contour classification and final template matching. This study focuses on recognizing five drug types: round tablets, oval pills, bicolor capsules, monochrome capsules, and cod liver oil. Recognition primarily relies on drug contour shapes and color features. Our solution first classifies drugs by contour features, then employs the SAD template matching algorithm for rapid identification of pharmacy-compliant drugs. This approach effectively improves recognition accuracy and speed, providing technical support for automated drug handling and transportation.Fig. 11Drug recognition workflow. Drug recognition workflow. Drug image preprocessing Image preprocessing plays a crucial role in drug recognition systems, aiming to enhance image quality, remove noise, and highlight key drug features for reliable subsequent recognition and classification. In drug identification, preprocessing mainly improves clarity of shape, color, and texture features while reducing interference from environmental factors and imaging devices. Image grayscale conversion First, images undergo grayscale conversion. Grayscale refers to color values in grayscale images ranging from 0 (black) to 255 (white), with intermediate values representing shades. Grayscale histograms statistically analyze pixel distributions for each grayscale value31. In drug recognition, grayscale reduces computational complexity by retaining only brightness information without color data, facilitating efficient edge detection and contour extraction. Common grayscale methods include maximum value, average value, and weighted average methods. The maximum method uses the highest RGB channel value, potentially causing excessive brightness. The average method calculates the mean RGB value but may lose color details. The weighted average method improves upon this by assigning different weights to RGB channels based on human color sensitivity:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Gray} = 0.299 \times R + 0.587 \times G + 0.114 \times B, \end{aligned}$$\end{document}where R, G, and B represent red, green, and blue channel values. These weights align with human eye sensitivity (higher green weight)32. We selected weighted averaging for better color retention and recognition accuracy. Image denoising Image denoising is a crucial step in image processing, aiming to eliminate noise introduced during image acquisition, transmission, or environmental interference, thereby restoring the true information of the image33. In pharmaceutical recognition systems, denoising is particularly critical, as noise can distort the shape and contour extraction of pharmaceutical products, ultimately degrading recognition accuracy. Noise introduces undesirable pixel variations, adversely affecting subsequent processing steps. Among linear filtering methods, Gaussian filtering stands out as an effective approach that overcomes the limitations of simpler neighborhood-based techniques, such as excessive blurring. Conventional methods like mean filtering, which compute a simple average of neighboring pixels, tend to blur fine details and edges. In contrast, Gaussian filtering employs a Gaussian distribution-based weighting scheme, assigning higher importance to nearby pixels during smoothing operations. By computing a weighted average where closer pixels contribute more significantly, Gaussian filtering achieves superior noise suppression while better preserving structural edges compared to median filtering (which is computationally intensive and may lose details under strong noise) and mean filtering (which often blurs structural contours). Consequently, Gaussian smoothing yields higher-quality results with improved edge retention, making it widely applicable in high-precision image processing scenarios. Given these advantages, this study adopts Gaussian filtering for denoising pharmaceutical images to ensure robust feature extraction and enhanced recognition performance. Image enhancement Image enhancement is a pivotal technique in image processing aimed at improving the visual quality of an image by adjusting its brightness, contrast, or detail characteristics, thereby making the salient information more prominent34. This process facilitates the clear presentation of target regions during image analysis, ultimately enhancing the accuracy and reliability of subsequent tasks, such as feature extraction or object recognition. In the field of pharmaceutical identification, image enhancement is widely employed, particularly in scenarios requiring high-quality image input to ensure analytical precision. Among the most commonly used techniques in image enhancement is grayscale transformation, which improves image contrast by adjusting the grayscale value of each pixel. Grayscale transformation can be categorized into linear and nonlinear methods. Nonlinear transformations, including exponential and logarithmic transformations, are designed to expand high- and low-grayscale regions, respectively. However, these approaches may introduce uneven illumination in cylindrical label imaging. In contrast, linear transformation effectively expands the dynamic range of an image, enhancing its original contrast and rendering it sharper, which significantly improves visual quality. Given input image f(x, y) with grayscale range M and output g(x, y) with range N, where input grayscale interval is [a, b] and output is [c, d], the piecewise linear transformation is:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g(x,y) = {\left\{ \begin{array}{ll} \frac{c}{a}f(x,y), & 0 \le f(x,y)< a \\ \frac{d-c}{b-a}[f(x,y)-a]+c, & a \le f(x,y) \le b \\ \frac{N-d}{M-b}[f(x,y)-b]+d, & b < f(x,y) \le M \end{array}\right. } \end{aligned}$$\end{document}Different a, b, c, d values yield varying effects:When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a = c,\ b = d$$\end{document}, the transformation is an identity function (slope=1).When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a>c,\ bc,\ bc,\ b 0.15) was determined through an extensive evaluation of classification performance across a range of candidate values. Systematic tests revealed a clear trade-off: an excessively low threshold (e.g. a value less than 0.10) increased the risk of misclassifying capsules with subtle surface textures as elliptical pills (lower recall), while an excessively high threshold (e.g. a value greater than 0.20) led to elliptical pills with simpler textures being falsely identified as capsules (lower precision). By comparing the recognition effects under different thresholds, it was found that the classification performance, optimally balancing precision and recall, was achieved at 0.15. Experiments prove that the LBP uniformity of capsules is mostly located in the range of 0.25–0.45, while that of elliptical tablets is approximately 0.05–0.20, with 0.15 providing a clear separation margin between the two distributions. Aspect ratio is a key feature for capsule identification. As shown in Fig. 12, circular tablets are concentrated in the range 0.9–1.1, fish oil capsules in 1.35–1.45, elliptical pills in 1.8–2.2 (centered at 1.9), and capsules in 2.0–2.3 (centered at 2.1). By setting an aspect ratio threshold of 2.0, most capsules are correctly identified, while circular tablets, fish oil, and the majority of elliptical pills remain below the threshold. In addition to shape, texture plays an important role. Local Binary Pattern (LBP) uniformity captures the regularity of surface patterns, where capsules generally exhibit higher uniformity values (0.25–0.45) and elliptical tablets lower values (0.05–0.20). Using a threshold of 0.15 for LBP uniformity, as illustrated in Fig. 13, further enhances discrimination, reducing misclassification and complementing the aspect ratio criterion. Together, these two features enable robust capsule recognition in the pharmacy dataset. This approach eliminates the need for complex fitting procedures, enhancing both efficiency and accuracy. This approach effectively distinguishes capsules from elliptical pills, eliminates the need for complex fitting procedures, and significantly improves both recognition efficiency and accuracy.Fig. 12Aspect ratio distribution of different pill types.Fig. 13LBP uniformity distribution of capsules and elliptical pills. Aspect ratio distribution of different pill types. LBP uniformity distribution of capsules and elliptical pills. Drug template matching Template matching is a method used to identify the presence of a known template image within a larger search image, despite variations caused by factors such as illumination and positional changes during image acquisition. By employing correlation-based search strategies, this approach locates sub-images similar to the template and determines their positions45. Conventional Sum of Absolute Differences (SAD)-based template matching often suffers from limited matching accuracy. To address this issue, this study proposes an enhanced SAD-based template matching algorithm integrated with contour feature classification, thereby improving both the efficiency and precision of pharmaceutical recognition. SAD-based template matching The Sum of Absolute Differences (SAD) is a widely used similarity metric in image matching, particularly for template matching and disparity computation in stereo vision. The fundamental principle of this algorithm lies in quantifying the similarity between two image regions by computing the differences in grayscale values of corresponding pixels within image patches. The computational process involves selecting fixed-size windows in both the reference and target images, followed by a pixel-wise comparison of grayscale values within these windows46. In practical implementation, a center point is first selected in the reference image, around which a fixed-size window is constructed. A corresponding sub-window of identical dimensions is then created at the same location in the target image, and the absolute differences between the grayscale values of corresponding pixels within these windows are computed. These differences are summed to yield a cumulative error value (SAD), which serves as a measure of dissimilarity between the reference and target windows. During the matching process, the algorithm traverses the entire reference image starting from the top-left corner, computing the SAD value for each candidate window relative to the corresponding region in the target image. By evaluating SAD values across different regions, the optimal matching position is identified as the location with the minimal SAD value, indicating the highest degree of similarity between the reference and target image regions. The formula is17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} C_{\text {SAD}} = \sum _{\{i,j \in W\}} |I_L(x+i,y+j) - I_R(x+i+d,y+j+d)|, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_L(x,y)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_R(x,y)$$\end{document} are left/right image pixel values, W is window size, and d is disparity. Although simple and efficient, SAD is noise-sensitive and affected by lighting changes47. The SAD algorithm offers notable advantages, including computational simplicity and high efficiency. However, its primary limitations lie in its sensitivity to noise and susceptibility to global illumination variations. To address these challenges, this study integrates contour feature analysis specifically edge feature detection to enhance the overall accuracy of SAD-based template matching in pharmaceutical product recognition. Contour-enhanced SAD matching Before SAD matching, drugs are pre-classified by geometric shape (circular, elliptical, capsule-shaped), reducing search space and improving efficiency. Within each class, further distinctions are needed (e.g. cod liver oil vs. oval pills in the elliptical class). Thus, SAD alone is insufficient without contour shape differentiation. To address the need for further distinctions within each shape class, we constructed dedicated template libraries through a standardized pipeline to ensure representativeness and matching accuracy. The library was built by first capturing images of the five target drug types under controlled lighting conditions using a fixed 5-megapixel industrial camera, yielding 300 raw images from multiple angles. Candidate templates were then manually selected from this pool, with approximately 50 images per category chosen for their clear contours, absence of occlusion, and typical features. To ensure that the template library adequately represents the variations in appearance due to changes in viewing angle and illumination, candidate templates were selected to include images captured under diverse orientations and lighting conditions. This selection strategy ensures that the SAD matching process remains robust against moderate rotations, tilts, and shadows commonly encountered in practical pharmacy environments. These candidate images subsequently underwent the same preprocessing pipeline detailed in Sect. "JPS-RRT* algorithm", including grayscale conversion, Gaussian denoising, and image enhancement, followed by uniform cropping and resizing to a standard 64x64-pixel dimension to normalize scale and brightness. The final template set for each category, comprising about 30-40 templates, was curated by performing an internal consistency check based on SAD values to remove outliers, thereby ensuring high library quality. This process resulted in a finalized template library comprising 130-140 templates per category (specifically: 135 for round tablets, 132 for oval pills, 130 for cod liver oil, 138 for monochrome capsules, and 140 for bicolor capsules). This specific scale was determined to be the optimal trade-off, balancing template diversity against computational load on our target embedded platform (Raspberry Pi). It ensures matching can be completed within a constrained time frame, demonstrating the method’s practicality and scalability for structured environments like intelligent pharmacies, where the drug types are predefined and stable. During the SAD matching process, only the templates from the corresponding contour category are invoked for comparison. This structured approach to library construction not only confines the search space to relevant templates, avoiding unnecessary computational overhead, but also significantly mitigates the risk of mismatches caused by fundamental shape discrepancies. In the SAD matching procedure, the input target image first undergoes preprocessing, including grayscale conversion, normalization, and size standardization. These steps ensure consistency between the input image and the templates in terms of scale, brightness, and other factors, thereby minimizing interference caused by variations in lighting or dimensions. Subsequently, based on the contour feature classification results obtained in the preliminary stage, the corresponding template library is selected. The SAD values between the target pharmaceutical product and each template in the library are computed, and the template with the smallest SAD value is identified as the matching result. This method eliminates the need for a global search, improving computational efficiency, while simultaneously excluding templates that do not belong to the current shape category during matching, thereby enhancing recognition accuracy48. To verify the effectiveness of the improved SAD template matching algorithm, this paper conducts a comparative analysis with the traditional SSD (Sum of Squared Differences) algorithm and the deep learning model YOLO. Both SSD and SAD are template-based matching methods, but SSD requires the calculation of the square of the difference for each pixel, resulting in a significantly higher computational load than SAD, which leads to longer processing time and higher requirements for processor performance25. Through comparative tests in a pharmacy environment, at the same resolution, the operation speed of the improved SAD algorithm in this paper is approximately 10-12% faster than that of the SSD algorithm, and the accuracy rates of both algorithms for drug recognition are close. YOLO is a typical deep learning-based object detection algorithm that transforms the object detection task into a single regression problem, simultaneously predicting the position and category of the target in a single forward pass. It features end-to-end processing and high-speed detection26. Compared with the YOLO deep learning method, the improved SAD algorithm in this paper has higher recognition accuracy, requires no training, and can run on embedded platforms such as Raspberry Pi with only the deployment environment, achieving good recognition accuracy under limited computing power. It combines lightweight and ease of use, making it suitable for low-power drug recognition tasks in pharmacy robots. Through the combined application of contour feature classification and SAD template matching, this method can significantly enhance the accuracy of pharmaceutical identification. This advancement provides reliable technical support for automated drug sorting, intelligent pharmaceutical management, and quality inspection (Table 10).Table 9Recognition accuracy comparison of the improved SAD algorithm and other methods.Matching algorithmRound tabletsOval pillsCod liver oilMonochrome capsulesBicolor capsulesSAD78.3%67.4%73.6%64.1%61.3%SSD79.7%68.6%74.9%65.3%61.9%YOLO97.7%95.4%93.7%95.3%95.8%Improved SAD91.6%86.2%88.4%82.4%85.8%Table 10Mismatch rate comparison of the improved SAD algorithm and other methods.Matching algorithmRound tabletsOval pillsCod liver oilMonochrome capsulesBicolor capsulesSAD21.7%32.6%26.4%35.9%38.7%SSD20.3%31.4%25.1%34.7%38.1%YOLO2.3%4.6%6.3%4.7%4.2%Improved SAD8.4%13.8%11.6%17.6%14.2%Table 11Matching time comparison of the improved SAD algorithm and other methods.Matching algorithmRound tabletsOval pillsCod liver oilMonochrome capsulesBicolor capsulesSAD9.34s12.57s10.21s13.12s12.71sSSD11.45s11.45s11.26s14.22s13.98sYOLO18.88s20.46s20.12s19.47s19.86sImproved SAD3.23s4.18s3.65s4.51s4.39sTable 12Average recognition accuracy and matching time for different algorithms.Matching algorithmAverage recognition accuracyAverage matching timeSAD68.94%11.59sSSD70.10%12.47sYOLO95.58%19.76sImproved SAD86.04%3.99s Table 9Recognition accuracy comparison of the improved SAD algorithm and other methods.Matching algorithmRound tabletsOval pillsCod liver oilMonochrome capsulesBicolor capsulesSAD78.3%67.4%73.6%64.1%61.3%SSD79.7%68.6%74.9%65.3%61.9%YOLO97.7%95.4%93.7%95.3%95.8%Improved SAD91.6%86.2%88.4%82.4%85.8% Recognition accuracy comparison of the improved SAD algorithm and other methods. Table 10Mismatch rate comparison of the improved SAD algorithm and other methods.Matching algorithmRound tabletsOval pillsCod liver oilMonochrome capsulesBicolor capsulesSAD21.7%32.6%26.4%35.9%38.7%SSD20.3%31.4%25.1%34.7%38.1%YOLO2.3%4.6%6.3%4.7%4.2%Improved SAD8.4%13.8%11.6%17.6%14.2% Mismatch rate comparison of the improved SAD algorithm and other methods. Table 11Matching time comparison of the improved SAD algorithm and other methods.Matching algorithmRound tabletsOval pillsCod liver oilMonochrome capsulesBicolor capsulesSAD9.34s12.57s10.21s13.12s12.71sSSD11.45s11.45s11.26s14.22s13.98sYOLO18.88s20.46s20.12s19.47s19.86sImproved SAD3.23s4.18s3.65s4.51s4.39s Matching time comparison of the improved SAD algorithm and other methods. Table 12Average recognition accuracy and matching time for different algorithms.Matching algorithmAverage recognition accuracyAverage matching timeSAD68.94%11.59sSSD70.10%12.47sYOLO95.58%19.76sImproved SAD86.04%3.99s Average recognition accuracy and matching time for different algorithms. Trade-offs and applicability of the improved SAD algorithm The results presented in Tables 7 and 8 reveal the inherent trade-offs between the improved SAD algorithm and the YOLO deep learning model across different performance metrics. Although the improved SAD exhibits a higher mismatch rate (17.6%) than YOLO (4.7%) for certain categories, such as monochrome capsules, this elevated error rate remains acceptable within the specific application context of an intelligent pharmacy’s unmanned delivery system. The fundamental reason lies in the system’s design philosophy, which adheres to an engineering trade-off criterion tailored for embedded platforms: prioritizing system-wide real-time performance, low power consumption, and overall reliability, while meeting basic accuracy requirements. The advantages of the improved SAD method become particularly prominent in resource-constrained scenarios. While the high accuracy of algorithms like YOLO relies on substantial computational load and parameter scale, necessitating GPU acceleration–feasible in high-power, high-cost server-side deployments–their real-time operation on resource-limited embedded edge devices like the Raspberry Pi is challenging. In contrast, the improved SAD algorithm requires no training, has a minimal memory footprint, and can operate efficiently in synergy with the JPS-RRT* path planning module on the same low-power computing platform.Specifically, deployment on a Raspberry Pi 4B (CPU: Quad-core Cortex-A72 1.5 GHz, RAM: 4GB) demonstrated an average processing time of 3–4.5 seconds per image, with CPU utilization around 40–50% and memory usage of 500–700 MB (Table 11). These measurements confirm that the algorithm can function effectively on low-power hardware without compromising system responsiveness. This integrated design avoids the degradation in real-time performance or system stability that could result from introducing a computationally intensive recognition model, thereby achieving optimal system-level performance. Furthermore, intelligent pharmacy environments typically possess structured characteristics such as relatively stable lighting, regular drug placement, and minimal background clutter, providing favorable conditions for the application of the improved SAD algorithm. Its principle, based on transparent template matching, allows for easy tracing and debugging of mismatch causes, significantly facilitating system maintenance and optimization. To provide a comprehensive evaluation of the proposed method, the recognition accuracy, mismatch rate, and matching time of different algorithms are summarized in Tables 9, 1011 and 12. As shown in Table 11, the improved SAD algorithm achieves significantly higher recognition accuracy than traditional SAD and SSD across all five drug categories. Correspondingly, Table 12 demonstrates that the mismatch rate also decreases notably, particularly for round tablets and oval pills. In terms of processing efficiency, Table 11 shows that the improved SAD algorithm offers the shortest matching time among all compared methods, completing recognition in only 3–4.5 seconds depending on the drug type. Furthermore, the overall comparison in Table 12 highlights that the improved SAD achieves the best balance between accuracy (86.04%) and efficiency (3.99 s), making it highly suitable for resource-limited embedded systems. To further clarify the quantitative indicators used in evaluating the improved SAD algorithm, we explicitly interpret how these metrics influence overall system performance. The mismatch rate of the traditional SAD method (27.33%) is reduced to 15.12% with the improved version, corresponding to roughly 60 fewer misidentifications in a 500-image dataset (500 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} (27.33% – 15.12%) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\approx$$\end{document} 61). This reduction directly lowers the risk of medication retrieval errors and enhances workflow stability in unmanned pharmacy operations. Although YOLO achieves higher accuracy (95.58%), its average processing time of 19.76 s per image introduces substantial latency for embedded robotic systems and may cause stalling or task backlogs during multi-drug retrieval tasks. In contrast, the improved SAD requires only 3.99 s per image, enabling recognition to run concurrently with the JPS–RRT* path-planning module on the same Raspberry Pi platform and preventing perception–planning bottlenecks. Thus, while the improved SAD sacrifices marginal accuracy compared to YOLO, its superior inference speed ensures continuous robot operation and higher overall throughput. These quantitative comparisons demonstrate that under embedded resource constraints, where real-time responsiveness, energy efficiency, and system-level stability outweigh small accuracy gains, the improved SAD algorithm achieves the most practical balance for intelligent pharmacy environments characterized by structured lighting, consistent drug placement, and long-term low-power operation requirements. In conclusion, the core competitiveness of the improved SAD algorithm lies not in pursuing ultimate recognition accuracy, but in achieving the best performance balance under specific constraints, including low power consumption, high real-time demands, and structured environments. It offers a practical and efficient solution for the unmanned delivery tasks in small to medium-sized intelligent pharmacies. For large-scale automated warehouses where extreme accuracy is paramount and hardware conditions permit, solutions based on high-performance hardware running deep learning algorithms could be considered. Conclusion This study presents a novel path optimization technique combining JPS and RRT* algorithms for unmanned delivery vehicles in smart pharmacy environments. Experimental results demonstrate that the JPS method’s pruning strategy significantly enhances global path search efficiency, while RRT* ensures path smoothness and dynamic environment adaptability through local optimization. Compared to conventional Dijkstra and A* algorithms, our approach shows superior performance in planning time (reduced by 82%), path quality (25% shorter average length), and computational efficiency (68% lower resource consumption), with enhanced dynamic environment adaptability. Complementing the navigation system, deep learning-optimized SLAM mapping improves positioning accuracy by 40% and environmental perception capabilities. For drug recognition, our multi-stage framework–incorporating template matching, shape classification, contour identification, and image preprocessing–achieves an average identification accuracy of 86.0% (see Table 12), enabling real-time pharmaceutical classification. Compared with the traditional SAD and SSD template matching algorithms, the improved SAD algorithm demonstrates significant improvements in both recognition accuracy and matching speed. Although it still lags behind the deep learning algorithm YOLO in terms of accuracy, the matching time required by the improved SAD is substantially shorter than that of YOLO. Validated in operational pharmacy settings, this integrated technology significantly improves medication distribution efficiency, reduces manual error rates by approximately 75%, and enhances workplace safety. The proposed solution not only optimizes smart pharmacy logistics but also provides a reference architecture for autonomous mobile robots operating in structured environments. Table 12 further illustrates that while YOLO achieves the highest accuracy, the improved SAD offers the best trade-off between recognition performance and computational efficiency compared with SAD, SSD, and YOLO.Future work will explore integrating reinforcement learning (RL) with the current system, where RL can dynamically optimize template selection, matching thresholds, and preprocessing parameters based on real-time feedback from recognition outcomes. By continuously learning from successes and failures in drug identification, the system can adapt to variations in lighting, occlusion, and placement of pharmaceuticals. This approach aims to improve overall recognition accuracy, reduce mismatch rates, and enable the system to generalize to new or unseen drug types, laying the foundation for more intelligent, autonomous, and robust drug handling in complex pharmacy environments. Since the JPS algorithm used as the global planning method has high requirements for the accuracy of the map, an accurate map needs to be drawn. If there is a certain error in the corner positions of the map, there is a probability that the vehicle will get stuck in a dead zone during the planning process, thereby causing a system crash. At the same time, when multiple dynamic obstacles appear simultaneously, although the algorithm in this project has reduced the computing power requirement, it is still found that the planning speed has decreased to some extent, resulting in a certain reduction in real-time performance. Therefore, this project is designed for scenarios where there are multiple static obstacles but also a small number of dynamic obstacles, and planning with low computing power can present more superior performance.Furthermore, the proposed drug recognition method has several inherent limitations that warrant consideration. First, its performance is highly sensitive to variations in lighting conditions; significant changes in illumination can affect the quality of contour extraction and reduce matching accuracy. Second, the method requires a relatively clean and uncluttered background to reliably detect and classify drugs. Background noise or overlapping visual elements may interfere with template matching and lead to misidentification. Third, when drugs are partially occluded or stacked on top of each other, the algorithm may struggle to correctly identify individual items, resulting in reduced recognition accuracy. These limitations indicate that while the method is effective in structured and controlled environments, its robustness may be challenged in more complex or dynamic scenarios. The primary limitation of the current system lies in its performance degradation when facing a high density of simultaneous dynamic obstacles, as the computational load for real-time replanning pushes the limits of the embedded Raspberry Pi platform. To concretely address this challenge, future work will focus on integrating existing, well-established cloud–edge computing paradigms. Specifically, we propose to offload the computationally intensive global path planning and multi-obstacle dynamic scene understanding tasks to the cloud using industrial solutions such as AWS IoT Greengrass or Azure IoT Edge. In addition to cloud–edge optimization, a significant future direction is the integration of reinforcement learning (RL) into the current JPS–RRT* framework. To clarify how this would be incorporated, we envision RL serving as a high-level adaptive controller that operates alongside the existing hybrid planner. Instead of replacing JPS or RRT*, the RL agent would learn to adjust key local planning parameters—such as the sampling bias \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {bias}}$$\end{document} or the obstacle inflation distance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$inflation\_dist$$\end{document}—based on real-time perception features (e.g. obstacle velocity, density, or predicted motion patterns). In this design, JPS continues to generate a reliable global path, RRT* ensures geometric feasibility for local motion, and the RL module dynamically tunes the planner’s behavior to improve reactivity and robustness in highly dynamic environments. This parameter-level integration offers a practical and computationally efficient pathway to enhance the current system while preserving its existing architecture.