A Bi-Level Hybrid Framework for Multi-Target Path Planning of AGV Based on Particle Swarm Optimization and Bidirectional Rapidly Exploring Random Tree Authors: Tursun Mamat, Zhaolong Liu, Qiuju Yang, Abdukeram Dolkun, Longfei Li Published: 2026 ABSTRACT Highlights What are the main findings? A bi-level collaborative framework integrating PSO and Bi-RRT is proposed for AGV multi-target path planning in complex logistics environments.Adaptive sampling, local-density suppression, and lightweight online parameter adjustment strategies are introduced to improve planning efficiency and reduce redundant exploration. What are the implications of the main findings? The proposed framework improves the balance between global planning efficiency and local path feasibility in cluttered environments.The method provides a practical solution for trajectory optimization and multi-target path planning of AGV in warehouse and logistics applications. AbstractMulti-target path planning for Automated Guided Vehicle (AGV) in complex logistics environments requires balancing planning efficiency, obstacle avoidance capability, and trajectory smoothness. To address these challenges, this paper proposes a bi-level collaborative framework integrating Particle Swarm Optimization (PSO) with the Bidirectional Rapidly Exploring Random Tree (Bi-RRT). The framework unifies adaptive sampling, online parameter optimization, and trajectory smoothing within a single planning architecture. Specifically, the framework constructs a five-dimensional particle encoding that includes the expansion step size and multi-level strategy switching thresholds. During the Bi-RRT expansion process, an expansion-failure-driven adaptive sampling mechanism is introduced to enhance search performance in cluttered environments, while local-density-based suppression and directional dispersion are employed to reduce redundant exploration. In addition, a lightweight PSO-based monitoring mechanism enables online adaptive parameter adjustment. For multi-target scheduling, a greedy heuristic based on a hybrid weighted graph determines the visitation sequence. Trajectory smoothness is further improved using cubic B-spline interpolation combined with bounded perturbation optimization. Experimental results demonstrate that the proposed framework improves planning efficiency while maintaining stable performance across environments with different obstacle densities. These results demonstrate the effectiveness of the proposed framework for multi-target AGV path planning in complex warehouse environments. 1. Introduction AGV is widely used in warehousing and logistics [1]. Path planning plays a key role in system efficiency and safety, aiming to generate smooth, collision-free trajectories from the start to the goal in complex environments. Similar requirements also arise in UAV swarms [2], Unmanned Surface Vehicles (USVs) [3], harvesting manipulators [4], surveying robots [5], and hull-maintenance robots [6], making path planning a fundamental component of autonomous navigation for various mobile platforms. Early studies mainly relied on graph-search and meta-heuristic algorithms. In graph-based methods, variants of the A algorithm, such as Q-learning-based DP-A [7] and RLHA* integrated with neural networks [8], improve heuristic functions and enhance adaptability in unknown environments. The GPTA* algorithm introduces global guidance to improve search performance in grid-based maps [9]. Meta-heuristic algorithms have also been widely applied to handle multi-objective optimization and complex constraints. Genetic Algorithm (GA) is commonly used for task scheduling [10], while PSO is applied to path conflict resolution [11]. Recent advances in swarm intelligence, such as hybrid multi-swarm PSO [12] and large-scale multi-swarm scheduling strategies [13], further enhance global search capability and robustness in complex optimization problems. Continuous ant colony optimization (ACOPA R) helps avoid local optima through random-walk behavior [14], and multi-agent reinforcement learning combined with GA has been used to solve the Traveling Salesman Problem (TSP) [15]. However, these methods often show limited computational efficiency in high-dimensional spaces and densely cluttered environments. Sampling-based Rapidly Exploring Random Tree (RRT) and its variants are widely used for path planning in high-dimensional and complex environments. However, these algorithms often exhibit slow convergence and unstable initial trajectories in narrow passages. Recent studies address these limitations by improving different components of the algorithm. At the sampling stage, Fast-RRT [16] and F-RRT [17] improve efficiency through refined parent-node selection, while SOF-RRT [18], FHQ-RRT* [19], Agile-RRT* [20], and GD-RRT* [21] reduce redundant exploration using spatial probability weighting, sparse sampling, or Gaussian-based methods. In addition, hybrid optimization strategies have been introduced to enhance AGV path planning performance, such as KGWO-based methods that integrate multiple optimization mechanisms [22]. During the search process, several methods adopt bidirectional expansion strategies. PSBi-RRT [23] introduces probability smoothing. APF-based methods guide the search direction using artificial potential fields [24], while goal-biased APF-RRT [25] further improves exploration efficiency. Hybrid-RRT [26] enhances the search process through hybrid sampling. Kinematic constraints are considered in several methods. KB-RRT [27] incorporates motion constraints into tree expansion. Methods designed for Ackermann steering and Dubins vehicles, such as RRTASV [28] and 3PDP [29], address path curvature continuity to support feasible vehicle motion. Recent studies have widely applied Reinforcement Learning (RL) to dynamic path planning. The expected mean gamma incremental RL [30] and the 2D LiDAR-based VFH-QL [31] incorporate state preference information to address multi-objective planning in environments without prior knowledge. Digital Twin (DT)-assisted federated RL [32] uses predictions from virtual environments to compensate for data synchronization delays in real-world systems. In addition, RL has been combined with geometric risk evaluation to address the scheduling of UAV edge computing nodes [33]. Existing methods, typically designed for constrained scenarios, often struggle to simultaneously balance search efficiency, trajectory smoothness, and operational safety for AGV navigating complex, dynamic industrial environments. Conventional algorithms are also prone to local optima in intricate terrains or generate paths with excessive inflection points, which degrade kinematic stability and increase energy consumption. This paper proposes a bi-level multi-objective path planning framework to address these limitations. While recent studies have explored methods such as Grey Wolf Optimization with A*, Conflict-Based Search, and Artificial Potential Fields for multi-target path planning, these conventional approaches often combine existing algorithms or rely on empirically fixed heuristics. In contrast, the proposed method establishes a dynamic coupling between high-dimensional improved PSO and adaptive Bi-RRT. The bi-level architecture treats parameter tuning as an online process by introducing a continuous feedback mechanism rather than treating it as an isolated offline task. At the upper level, PSO performs dynamic optimization over a five-dimensional hyperparameter set using dimensional decoupling and stagnation-aware updates. The resulting parameters continuously guide the lower-level Bi-RRT, which adopts a multi-level sampling strategy driven by expansion failures to improve exploration efficiency and escape local optima. Based on the generated geometric paths, trajectory smoothness is further enhanced using a range-jitter-constrained PSO combined with cubic B-splines, ensuring kinematic feasibility. Finally, a mixed-weight graph model is employed to determine the visiting sequence for multi-target tasks. The proposed framework integrates path generation, smoothing, and task sequencing within a unified structure, demonstrating strong adaptability and improved performance compared with conventional hybrid planners. Table 1 summarizes the novelty of the proposed framework in contrast to the aforementioned recent methods. 2. Problem Formulation and Environment Modeling 2.1. Problem Statement and Assumptions To analyze the efficiency and smoothness of the proposed bi-level collaborative path planning mechanism, the operational environment of a large-scale logistics center is simplified. The problem formulation is based on the following assumptions: Single-AGV Setting: The study considers multi-target path planning for a single AGV. Multi-agent coordination and complex scheduling interactions are not included. Static Environment: The warehouse is modeled as static. Dynamic elements commonly present in real systems—such as moving obstacles, other vehicles, and temporarily placed goods—are not considered. Point-Mass Model: The AGV is represented as a point mass. Vehicle dynamics and kinematic constraints, including inertia, non-holonomic motion, and minimum turning radius, are not explicitly modeled. The focus is placed on geometric path-planning performance. Unrestricted Free Space: The AGV can move freely within the free space Ωfree, provided that the corresponding grid cell is unoccupied. The characteristics of the considered problem are summarized as follows. The inputs include the grid-based environment map S(x,y), the AGV starting position v0, and the predefined set of target nodes V. The outputs consist of the target visiting sequence π and the corresponding collision-free B-spline trajectory C(u). The optimization objectives are to minimize the global multi-target traversal cost J(π), reduce local path search time, and improve trajectory smoothness while maintaining collision-free navigation. 2.2. Grid-Based Environment Representation The warehouse environment is represented using a grid-based model. The workspace is defined as a two-dimensional domain Ω and discretized into an M × M uniform grid. The coordinate origin is located at the lower-left corner, and each grid cell has a unit side length. Environmental information is encoded using Boolean values. A state function S(x,y)∈{0,1} is defined, where S(x,y)=1 denotes an occupied cell belonging to the obstacle region Ωobs, and S(x,y)=0 indicates that the cell lies in the free space Ωfree. Since the AGV is modeled as a point mass, obstacle clearance is implicitly handled by the grid structure. No additional safety margin is incorporated into the distance cost function. Some grid cells are designated as task target points, shown as blue markers in Figure 1. 3. Improved PSO-Bi-RRT Bi-Level Collaborative Algorithm Traditional adaptive RRT methods often rely on static goal biasing or distance-based step-size adjustment. The improved Bi-RRT in this study introduces a closed-loop feedback mechanism that integrates failure-driven state transitions, density-based dynamic repulsion, angular constraints, and a continuous expansion strategy. This mechanism enables the algorithm to adapt its search behavior to local geometric features rather than relying on purely random exploration. 3.1. Improved Bi-RRT Algorithm (1)Multi-level sampling strategy driven by expansion failure count Multi-level sampling strategy driven by expansion failure count Existing RRT methods often rely on fixed empirical probability distributions for sampling. This method introduces a multi-level state transition mechanism. A dynamic multi-level adaptive sampling strategy based on the expansion failure count Nfail is proposed to improve the expansion efficiency of the random tree in complex environments. The reference direction θref is defined from the node Xnear, closest to the goal, toward Xgoal. The algorithm is divided into the following four adaptive stages: (1) Goal-directed Stage (Nfail≤2): In collision-free regions along the reference direction, a goal-directed greedy strategy accelerates tree expansion. The goal node Xgoal can be selected as the sampling point, with sampling also performed within a small neighborhood around θref, promoting rapid convergence toward the goal. (2) Locally directed expansion Stage (2Nstage2): When all preceding expansion attempts fail, the current node is considered to be trapped in a deep local minimum, such as a U-shaped configuration. Sampling is extended to the full 360° space. The sampling distribution is biased toward the rear region to promote backtracking and escape from spatial traps; a probability of 10% is assigned to the forward ±80° region, 20% to the regions between ±80° and ±120°, and the remaining 70% to the regions between ±120° and ±180°. Preliminary experiments identified Nstage1∈[3,8] and Nstage2∈[3,8] as suitable optimization ranges. Values below 3 tend to trigger excessive strategy switching, whereas values above 15 are frequently associated with prolonged search stagnation in highly constrained regions, such as U-shaped obstacles. The upper-level PSO performs online optimization within these ranges, reducing the reliance on manually selected threshold values. Figure 2 illustrates the multi-level adaptive sampling strategy guided by the number of expansion failures: (2)Dynamic Invalid-Radius Suppression Mechanism Based on Local Density Dynamic Invalid-Radius Suppression Mechanism Based on Local Density A newly generated candidate node Xnew is accepted if its distance to any existing tree node Xi satisfies the condition min∥Xnew −Xi ∥≥Rdyn. In this context, the adaptive invalid radius Rdyn is defined as:(1)Rdyn(Xnew)=Rbase⋅(1+η⋅NlocalNtotal)(2)Rbase=S ⋅ κ where Rbase denotes the base invalid radius, S represents the expansion step size, and κ∈[0.2,0.7] is the invalid-radius coefficient. Additionally, Ntotal indicates the total number of nodes in the tree, Nlocal corresponds to the number of nodes within a local neighborhood centered at Xnew with radius Rbase, and η∈[3,8] is the density penalty factor. In sparse regions, Rdyn≈Rbase, allowing the algorithm to perform fine-grained exploration. In contrast, in high-density regions, Rdyn increases, and candidate nodes falling within the invalid radius are excluded from connection. (3)Reference-Angle Dispersion Optimization Reference-Angle Dispersion Optimization Existing Bi-RRT algorithms typically rely solely on Euclidean distance to maintain node diversity and rarely consider the angular distribution of branches, which tends to produce redundant parallel paths. To address the issue, this study introduces an angle-repulsion constraint to enhance the diversity of search directions. The algorithm maintains a set of directions, Φ={ϕ1,ϕ2,...,ϕn,}, corresponding to existing branches within the neighborhood of the nearest node Xnear. For each newly generated candidate sampling angle, the minimum angular distance to any direction in the set Φ is computed as:(3)Δθmin=min|θrand −ϕj | ,(ϕj∈Φ) where Δθmin denotes the current minimum angular, δth=25° is the angle-repulsion threshold, and θrand represents the candidate sampling angle. If Δθmin<δth, the direction is considered overly congested, and θrand is forcibly reset. (4)Continuous Expansion Based on Hysteresis Decision and Directional Inertia Continuous Expansion Based on Hysteresis Decision and Directional Inertia Conventional Bi-RRT relies on single-step expansion, which is prone to path oscillations in narrow passages. This strategy introduces a continuous expansion mechanism to accelerate tree growth and improve obstacle avoidance stability. (1) To utilize directional inertia for accelerating the growth of the random tree, when the algorithm is in the obstacle-avoidance stage (Nfail≥Nstage1) and a new node Xnew is successfully generated, a greedy probing expansion is performed five times along the current sampling direction. If any candidate node encounters a collision or exceeds the boundary, the inertia-based expansion is immediately terminated. (2) Single-step sampling may be insufficient for escaping obstacle regions. A hysteresis decision mechanism is introduced to maintain the current sampling strategy until a stable exit is achieved. After the inertia-based expansion, a multi-step straight-line collision check is performed from the new node toward the global goal Xgoal. If the path is collision-free, the branch of the random tree is considered to have exited the obstacle, Nfail is reset to zero, and the algorithm returns to the direct-steering mode. Otherwise, Nfail remains unchanged, and the lateral detour mode is maintained in the next iteration. 3.2. Multi-Strategy Improved PSO Algorithm Due to the high coupling between parameters such as the expansion step size and the sampling-angle constraint threshold, the presence of multiple local optima in the solution space, and the strong stochastic noise inherent in the evaluation process, conventional PSO is prone to premature convergence in the context of Bi-RRT algorithms. To address these limitations, this study proposes an improved PSO algorithm that integrates high-dimensional particle encoding, logistic mapping, a comprehensive CLS mechanism, and a stagnation-based reset strategy. Figure 3 (Flowchart of the Multi-Strategy Improved PSO Algorithm) shows a detailed flowchart of the proposed multi-strategy improved PSO algorithm. Colored highlights indicate the components that contribute to the improvement of the search process. Furthermore, a linearly decreasing scheme is employed to dynamically adjust the inertia weight throughout the optimization process.(4)ω(t)=ωmax−tTmax(ωmax−ωmin) where Tmax denotes the maximum number of iterations, ωmax=0.9, ωmin=0.4, and t is the current iteration number. (1)High-Dimensional Particle Encoding Structure High-Dimensional Particle Encoding Structure The position vector of the i-th particle, Xi, is defined as a five-dimensional hyperparameter set encompassing three categories of parameters: motion characteristics, decision thresholds, and spatial constraints.(5)Xi=[S,Nstage1,Nstage2,κ,η] where S∈[0.5,1.5] denotes the base expansion step size. The state transition thresholds Nstage1 and Nstage2 correspond to the conditions governing the transitions within the multi-level sampling strategy. The dynamic suppression parameters κ and η regulate the effective range of the density-based adaptive invalid radius Rdyn, where κ represents the base invalid-radius coefficient, and η denotes the density-penalty strength. The complete list of optimized parameters is summarized in Table 2. (2)Particle Population Diversity Enhancement Particle Population Diversity Enhancement The logistic chaotic map is employed to enhance the diversity of the particle population within the five-dimensional hyperparameter search space. The velocity update formula utilizes a CLS. CLPSO enables each particle, when updating its velocity across different dimensions, to learn from the individual historical bests of distinct particles in the population, without relying on global best guidance. This dimensional decoupling mechanism helps prevent the swarm from premature convergence to local optima. (3)Stagnation-Detection-Based Population Diversity Maintenance Stagnation-Detection-Based Population Diversity Maintenance Each particle is equipped with a stagnation counter, τi. During each iteration, when updating pbest, if the fitness value of particle i changes, the counter τi is reset to zero; otherwise, it is incremented by one. Once τi exceeds the predefined threshold τlimit, the particle is considered to have become trapped in a local optimum or to have lost its search capability. The following remedial actions are then applied to the particle to restore population entropy and guide the algorithm out of local stagnation: 1. Position Reset: The particle’s position is randomly regenerated within the solution space. 2. Kinetic Energy Injection: The particle is assigned a relatively large random initial velocity. 3.3. Full-Process Collaborative Mechanism of Improved PSO and Bi-RRT While Bi-RRT provides efficient exploration in high-dimensional spaces, its performance is sensitive to parameter settings and may suffer from local search stagnation in complex warehouse environments. PSO, in contrast, offers strong global search capability and adaptive optimization characteristics. The proposed two-layer collaborative framework integrates PSO with Bi-RRT, enabling dynamic parameter adjustment during planning and alleviating search stagnation in complex environments. The overall workflow of the framework is illustrated in Figure 4. 3.3.1. Definition of the Fitness Function To mitigate the influence of stochasticity in the Bi-RRT algorithm on parameter optimization, a mean-based evaluation strategy is adopted during the fitness assessment stage. For each parameter combination Xi represented by a particle, K independent Bi-RRT path planning trials are executed consecutively, and the average path cost, average computation time, and success rate are computed. The fitness function is then formulated as follows:(6)Fitness(Xi) = α ⋅ T¯search + β ⋅ L¯path + γ ⋅ N¯collision + Pfailure where α, β, and γ are weighting coefficients assigned values of 0.4, 0.4, and 0.2, respectively. T¯search, L¯path, and N¯collision denote the average path planning time, path length, and number of collisions over K Bi-RRT trials, respectively, while Pfailure represents a penalty term. To balance evaluation accuracy and computational efficiency, two evaluation modes are defined: 1. Offline Evaluation Mode: A mean-based strategy with K=3 trials is employed. 2. Online Evaluation Mode: A fast evaluation strategy with K=1 is adopted. 3.3.2. Offline and Online Lightweight Improved PSO Mechanism (1)Offline Pre-optimization of Environment-Adaptive Parameters Offline Pre-optimization of Environment-Adaptive Parameters Prior to the execution of the improved Bi-RRT algorithm, the PSO algorithm is first initialized to perform offline optimization of the parameter set [S,Nstage1,Nstage2,κ,η], based on the obstacle-density characteristics of the current warehouse map. This process aims to identify an optimal combination of the expansion step size, strategy-switching thresholds, and density suppression coefficients in advance. (2)Online Hybrid Trigger Mechanism and Lightweight Fine-tuning Online Hybrid Trigger Mechanism and Lightweight Fine-tuning During the Bi-RRT expansion process, a hybrid monitoring mechanism is designed to continuously assess the search state and detect potential deadlocks. When the following condition is satisfied, the PSO-based optimization of Bi-RRT is triggered: (1) Excessive Collision Rate: If the proportion of collision failures ρcoll exceeds a predefined threshold within 25 consecutive expansion attempts, the PSO mechanism is activated.(7)ρcollision=Ncollision25 ⋅100%>70% where ρcollision denotes the collision rate and Ncollision represents the number of collision events. Based on extensive simulation-based calibration, the collision-rate threshold is set to 70%. (2) Expansion Stagnation: The growth status of the search trees is continuously monitored. If, over 50 consecutive iterations, the node growth rates, pgrow, of both trees fall below a predefined threshold, the algorithm is considered to have entered a stagnation state:(8)pgrow=Ntotal(t) − Ntotal(t−50)Ntotal(t−50)⋅100%<5% where pgrow denotes the node growth rate and Ntotal(t) represents the total number of nodes in the tree at iteration t. Based on extensive simulation-based calibration, the node growth rate threshold is set to 5%. (3) Time-Limit Constraint Early Warning: If the current iteration exceeds 60% of the maximum allowed iterations and the two trees have not yet connected, the algorithm is considered to be potentially trapped in a local optimum. 3.4. Path Post-Processing and Trajectory Smoothness Optimization This study proposes a PSO-based trajectory collaborative optimization strategy with range-jitter constraints. Redundant nodes in the initial path are first removed using a line-of-sight method, and the resulting path nodes are mapped to the initial control points of a B-spline curve. Subsequently, the PSO algorithm performs fine-tuning of the control points within their respective searchable regions to identify the optimal set of B-spline control points. 1. Jitter Space Construction and Particle Encoding Let the set of path nodes after redundancy elimination be denoted as P={p1,...,pn}. Except for the start and end nodes, which are fixed, the searchable region of each intermediate node pi is defined as a circular domain Ωi centered at its original position:(9)Ωi={(x,y)∈R2∣‖(x,y)−pi‖≤Rjitter} where Rjitter denotes the radius of the searchable domain for each node. Rjitter is set to 0.3S, where S is the Bi-RRT expansion step size. This bounded-jitter constraint limits excessive deviation from the original feasible path while preserving sufficient optimization flexibility. If a control point enters an obstacle region during particle updates, the corresponding particle is either projected onto the boundary of the feasible domain or reset to its last valid position. Furthermore, after each control-point update, the resulting B-spline trajectory is uniformly discretized and evaluated for collisions. Trajectories that intersect obstacles are assigned a large penalty during fitness evaluation, preventing them from being selected as feasible solutions. Consequently, the optimization process is restricted to collision-free trajectories, ensuring obstacle avoidance throughout trajectory generation. This optimization process is illustrated in Figure 5. 2. Collaborative Fitness Evaluation The fitness function Fsmooth directly evaluates the B-spline trajectory C(u) induced by the current control points, thereby jointly optimizing control point placement and trajectory quality. The evaluation integrates path length, smoothness, and safety:(10)Fsmooth=wlen⋅LC+wsmooth⋅∑k(1−vk⋅vk+1)2+PobsC where L(C) denotes the arc length of the generated curve, and Pobs(C) represents a collision penalty term defined via a penalty function. This term is assigned a sufficiently large value, for example, 1010, when any sampling point collides with an obstacle. vk denotes the unit tangent vector at sampling point k, where k indexes the discretized curve. The weights wlen and wsmooth correspond to the curve length and smoothness terms, respectively, with wsmooth≫wlen; in this study, they are set to wsmooth=50 and wlen=1. 3.5. Multi-Target Visit Sequence Planning Based on Hybrid Graph Weights A complete graph model with hybrid edge weights is constructed, and a greedy algorithm is employed to determine an AGV operation sequence that minimizes both energy consumption and time cost. (1)Problem Modeling and Hybrid Weight Strategy Problem Modeling and Hybrid Weight Strategy The AGV task scheduling problem is formulated as a complete undirected weighted graph G=(V,E). The vertex set V={v0,v1,…,vN} consists of the starting node v0 and N-1 target locations to be visited. A two-level weighting scheme is adopted for the edge weights Wij. If the straight-line path between nodes vi and vj is collision-free, Wij is defined as the Euclidean distance between them. Otherwise, an obstacle-avoiding path is generated using the improved PSO and Bi-RRT algorithm, and the corresponding path length is assigned to Wij. If path planning fails, Wij is set to infinity. (2)Construction of the Comprehensive Obstacle-Avoidance Cost Matrix Construction of the Comprehensive Obstacle-Avoidance Cost Matrix An environment complexity penalty factor is introduced to adjust the physical path length, thereby constructing a composite cost matrix C that reflects traversal cost. For non-line-of-sight paths, the matrix elements are defined as follows:(11)Cij=Lij+cbase⋅(α+β)(12)cbase=Lij−dij(13)dij=∥vi−vj∥2(14)α=1+ρ(15)ρ=Nobs/Ntotal(16)β=1+0.1⋅Npoint where Cij denotes the element of the composite cost matrix, and Lij represents the length of the collision-free obstacle-avoiding path between nodes vi and vj. The cbase denotes the baseline detour cost incurred by obstacle avoidance. The Euclidean distance between the two nodes is denoted by dij. The coefficient α is the obstacle density factor, where Ntotal denotes the total number of grid cells along the straight-line path between two nodes, and Nobs represents the number of obstacle-occupied grid cells among them. The parameter β is the path complexity coefficient, and Npoint denotes the number of turning points along the path where the turning angle satisfies Δθ>30°. (3)Greedy-Based Visit Sequence Optimization Greedy-Based Visit Sequence Optimization Based on the cost matrix C, the multi-target planning problem is formulated as an optimization over permutations π=(π0,π1,…,πN), with the objective of minimizing the global operational cost function J(π):(17)minJ(π)=∑k=0N−1Cπk,πk+1 The solution is subject to the constraint π0=v0, and {π1,…,πN}, which forms a permutation of the target set. Considering that the preceding Bi-RRT stage has already consumed part of the computational resources and that real-time performance is critical in warehouse scheduling, a nearest-neighbor greedy strategy is adopted to obtain a near-optimal sequence. The time complexity of this strategy is O(N2), which enables millisecond-level responsiveness for multi-target tasks. 4. Experiment 4.1. Simulation Platform and Parameter Settings Simulation experiments are conducted in MATLAB R2021a to evaluate the effectiveness and robustness of the proposed algorithm in complex warehouse environments. All experiments are performed on a laptop equipped with an Intel Core i7-13700H processor and 16 GB of memory. The proposed framework is compared with three conventional algorithms (Bi-RRT, RRT, and RRT-Connect), a Bi-RRT variant with offline PSO, and an improved Bi-RRT without PSO. In addition, two representative heuristic-guided methods, namely Artificial Potential Field-guided Bi-RRT (APF-Bi-RRT) and Grey Wolf Optimizer-guided Bi-RRT (GWO-Bi-RRT), are included to evaluate the effectiveness of different hybrid guidance mechanisms. All evaluated algorithms are run under the same computational budget and identical hardware configurations to ensure a fair and reproducible comparison. A unified set of fundamental parameters is applied across all methods. The expansion step size is set within the range [0.5, 1.5], and the connection radius is defined to be equal to the step size. The maximum number of iterations is set to 4000, 5000, 10,000, 15,000, and 15,000 for the five test scenarios, respectively. A consistent stopping criterion is applied: each algorithm terminates upon finding a valid collision-free path between the start and goal or upon reaching the maximum number of iterations, in which case the run is recorded as a failure. Simulations are conducted on three maps with different obstacle scales, along with two constrained scenarios, namely a U-shaped obstacle and a narrow passage. Performance is evaluated in terms of path length, sampled node count, and computation time. A statistical analysis is conducted to evaluate performance improvements and account for the stochastic nature of sampling-based planners. All algorithms are independently executed 50 times under identical configurations. Results are reported as mean ± standard deviation (SD) to reflect performance variability and stability. Given that RRT-based results may exhibit non-normal distributions and heteroscedasticity, a non-parametric Wilcoxon rank-sum test is adopted instead of a parametric test. This test assesses the significance of performance differences between the proposed method and baseline algorithms. For all comparisons, the results yield p < 0.01, indicating statistically significant performance differences. 4.2. Comparison of Path Planning Trajectories Among Different Algorithms in Complex Warehousing Environments Environment A includes three obstacle scenarios: sparse, moderate, and dense. Five target nodes are specified, resulting in a total of ten undirected test paths. For visual clarity, only the planning process for the node pair with the maximum Euclidean distance is illustrated, while evaluation is performed using data from all ten paths. For completeness, detailed trajectory visualizations for all independent planning instances in the representative 44 × 44 dense obstacle scenario are provided in the Supplementary Materials (Figures S1–S9 and Table S1). The expansion of the start and goal trees is represented by green and blue lines, respectively. The three environments are shown in Figure 6, where blue markers indicate the target nodes. The proposed algorithm combines dynamic sampling strategies with online PSO refinement, enabling the random tree to escape local minima and reducing ineffective sampling. The simulation trajectories of the various compared algorithms in the first scenario (the 22 × 22 sparse obstacle environment) are shown in Figure 7, and the corresponding parameter statistics curves are summarized in Figure 8. Furthermore, the simulation maps and statistical data for the second scenario (the 33 × 33 medium obstacle environment) are presented in Figure 9 and Figure 10. Finally, the path planning trajectories and data statistics for the third scenario (the 44 × 44 dense obstacle environment) are illustrated in Figure 11 and Figure 12. Experimental results show that the proposed method outperforms the benchmark algorithms in terms of average path length, sampled node count, and computation time. As environmental complexity increases, the proposed algorithm achieves more significant improvements in computational efficiency and obstacle avoidance. The aggregated results over all planning instances and their corresponding instance-level evaluations are summarized sequentially according to the environmental complexity: Table 3 and Table 4 for the 22 × 22 scenario, Table 5 and Table 6 for the 33 × 33 scenario, and Table 7 and Table 8 for the 44 × 44 scenario. The instance-based analyses exhibit trends consistent with the aggregated statistics, indicating that the advantages of the proposed framework are not confined to a limited number of favorable cases. Despite performance variations caused by different obstacle configurations and planning tasks, the proposed method consistently demonstrates superior planning efficiency and reduced sampling requirements, thereby confirming its robustness and effectiveness across diverse environments. The introduced Logistic chaotic mapping strategy enhances population diversity, leading to faster convergence of the fitness curve (Figure 13a). Meanwhile, the PSO mechanism iteratively adjusts parameters to obtain an effective configuration for the current environment. When tree expansion stagnates in complex obstacle regions, an online monitoring mechanism activates a PSO-based adjustment to guide the search. This helps the algorithm escape local minima and traverse constrained regions (Figure 13b,c). In extreme environments, the proposed algorithm produces fewer redundant expansions than the baseline methods (Figure 14 and Figure 15). The quantitative simulation results for the narrow passage scenario (Scenario B) are summarized in Table 9. In the U-shaped obstacle scenario (Scenario C), although RRT* occasionally yields shorter paths, as shown in Table 10, this is achieved at the cost of a higher number of sampled nodes. Overall results indicate that the proposed method outperforms the benchmark algorithms, particularly in reducing the average computation time and sampled node count. 4.3. Path Post-Processing and Multi-Target Visit Sequence Validation The effectiveness of the trajectory-smoothing strategy is evaluated using paths generated in the three configurations of Environment A (Figure 16). Under PSO-based bounded-jitter constraints, the cooperative B-spline optimization removes abrupt changes in piecewise-linear segments, producing a continuous and smooth curve. Compared with the original trajectory and the pruned path, the optimized trajectory exhibits improved smoothness. As shown in Figure 17, the optimal visiting sequence of the AGV in Environment A is given as follows: Start“→” Target4“→” Target1“→” Target2“→” Target3; Start“→” Target3“→” Target4“→” Target2“→” Target1; Start“→” Target2“→” Target3“→” Target4“→” Target1. 5. Conclusions This paper proposes a multi-objective path planning algorithm that integrates an improved PSO with an adaptive Bi-RRT to address the limitations of planning efficiency and obstacle avoidance in complex, high-density warehouse environments. Simulation results under different levels of complexity lead to the following conclusions: 1. This paper proposes a five-dimensional particle encoding scheme for kinematic features and decision thresholds and integrates PSO with Logistic chaotic mapping and CLS to achieve adaptive parameter tuning, improving the adaptability of the path planning algorithm across different environments. 2. This paper proposes an expansion strategy that integrates multi-level adaptive sampling, local density constraints, and orientation dispersion to adjust sampling directions and node distributions. Compared with conventional random expansion, the proposed method improves tree growth and alleviates stagnation during expansion. In addition, an online detection mechanism combined with a lightweight PSO supports real-time parameter tuning, improving the adaptability of the path planning framework in complex environments. Furthermore, statistical analysis indicates that the combination of the online detection mechanism and the lightweight PSO reduces search time and the number of sampled nodes. The proposed method exhibits lower SD and achieves statistical significance compared with all baseline algorithms, suggesting improved efficiency and stability. 3. This approach generates smooth, collision-free trajectories by combining a range-jitter-constrained PSO with cubic B-splines and obtains a cost-optimal visiting sequence using a greedy strategy on a mixed-weight graph. Despite these results, the proposed method exhibits several limitations. First, although lightweight PSO enables online parameter tuning, the iterative evaluation of particle fitness introduces additional computational overhead, which may affect performance in large-scale grid maps. Second, the current framework assumes a static global environment and relies on prior map information. Future work should consider the application of AGVs in dynamic obstacle scenarios that are frequently encountered in real-world operations. Specifically, introducing dynamic RRT variants could significantly improve real-time obstacle avoidance capabilities. Further research will explore coordinated collision avoidance in multi-AGV systems by incorporating temporal conflict detection mechanisms to enhance autonomous operation in complex warehouse environments. In addition, greater attention should be given to employing multi-objective optimization to balance path length, computational efficiency, and safety when improving local path planning during collision avoidance.