A Dual-Layer Hybrid-A* Path Planning Algorithm for Unstructured Environments Based on Phase Windows Authors: Tianxiao Zhu, Ziyu Xu, Rujiang Zhu, Wei Zhang, Zhonghua Miao Published: 2025 ABSTRACT In mobile robotics, path planning enables autonomous navigation to specified destinations. However, complex terrain can lead to excessive tilting or even overturning, compromising stability and safety. Traditional path-planning algorithms often fail to fully account for dynamic terrain variations and robot motion constraints. To address these limitations, this paper proposes the novel dual-layer Hybrid-A* algorithm, enhanced with dynamic phase windows. This approach represents a significant innovation by integrating real-time feedback mechanisms and adaptive adjustments to phase windows, enabling continuous path refinement in response to both environmental changes and robot motion limitations. The guidance layer introduces a bicubic interpolation-based super-resolution technique to refine elevation maps, offering more accurate posture estimation. In the planning layer, we propose the dynamic use of multiple cost functions, an adaptive expansion radius, pruning strategies, and a phase-window activation mechanism, effectively addressing the computational challenges posed by large search spaces. The integration of these strategies allows the algorithm to outperform traditional methods, particularly in unstructured environments with complex terrain. Experimental results demonstrate the effectiveness of the proposed method in generating optimized paths that satisfy robot motion constraints, ensuring both efficiency and safety in real-world applications. 1. Introduction Mobile robots have been extensively deployed in domains such as logistics, agriculture, defense, and aerospace, where they increasingly replace humans in repetitive, hazardous, or heavy-load tasks [1,2]. Global path planning enables mobile robots to autonomously construct feasible operational routes based on global environmental maps [3,4]. Consequently, path-planning approaches must address two fundamental challenges: generating accurate environmental representations and performing independent path searches within these maps [5]. These challenges directly influence the operational performance and efficiency of mobile robots. In general, a feasible planned path is expected to satisfy, either fully or partially, the following four properties: the adherence to the robot’s motion constraints, conformity with environmental constraints, completeness of the solution, and optimality of the generated trajectory. Achieving these properties typically requires a balance between planning quality and computational efficiency, as improving one aspect often comes at the expense of the other [6]. Common path-planning algorithms typically operate on binary maps, where environmental constraints are simplified to passable or non-passable regions, and the mobile robot is treated as a point mass [7,8,9,10]. These approaches perform well in structured environments such as warehouses and logistics facilities. However, in real-world scenarios, three-dimensional terrain significantly influences both the traversability and the cost of robot motion [11]. As a result, optimizing paths based solely on two-dimensional projections proves insufficient and often inaccurate. Three-dimensional path planning methods are commonly employed in UAV applications, where 3D cells or polygonal representations are used to simplify the planning process [12,13]. Nevertheless, because UAVs can freely maneuver along the vertical axis, the assumptions underlying their planning algorithms differ substantially from those of ground mobile robots. As a result, UAV-oriented methods cannot be directly applied to terrestrial robots operating under strict kinematic and terrain constraints. While concepts such as phase windows, coarse guidance, and Dubins-based regions have been widely applied in path planning, these techniques are often limited to static environments or initial coarse path generation. Existing methods, such as A* and RRT, utilize these components to refine paths, but they fail to fully account for real-time environmental dynamics, kinematic constraints, or dynamic terrain variations. This approach addresses these gaps by integrating dynamic phase windows and real-time feedback mechanisms, allowing for continuous refinement of paths based on environmental feedback and robot kinematics. The posture of ground-operating mobile robots varies with the undulations of the terrain. In unstructured environments, in particular, posture variations can significantly affect the robot’s operational capability and stability [14]. To ensure safe and reliable autonomous operation, path planning for mobile robots must therefore incorporate comprehensive environmental and kinematic constraints. As a result, the cost function for path planning in unstructured environments should account not only for path length but also for factors such as terrain slope, vibration intensity, and the minimum turning radius of the robot. To address these requirements, this paper proposes a novel dual-layer Hybrid-A* algorithm that integrates dynamic phase windows and operates on DEM elevation maps of unstructured environments. Unlike traditional methods that rely on static phase windows and coarse guidance, this approach introduces a dynamic phase-window adjustment strategy, enabling real-time path refinement based on environmental feedback and kinematic constraints. This dynamic adjustment ensures continuous optimization of paths, adapting to both environmental changes and robot motion limitations. The guidance layer provides landmark guidance points and a phase-window partitioning strategy, while the planning layer implements phase-window switching mechanisms and an enhanced Hybrid-A* method that accounts for both environmental and kinematic constraints. A dual-layer Hybrid-A* path planning framework is developed to effectively reduce the search space size and improve overall path search efficiency. A dual-layer Hybrid-A* path planning framework is developed to effectively reduce the search space size and improve overall path search efficiency. A bicubic interpolation–based super-resolution reconstruction method is introduced to refine elevation maps in unstructured environments, providing a reliable basis for mobile robot posture estimation in the planning layer. A bicubic interpolation–based super-resolution reconstruction method is introduced to refine elevation maps in unstructured environments, providing a reliable basis for mobile robot posture estimation in the planning layer. Phase windows constructed in the guidance layer, along with their associated heuristic cost functions, efficiently constrain the search region and guide the search direction. Phase windows constructed in the guidance layer, along with their associated heuristic cost functions, efficiently constrain the search region and guide the search direction. An improved Hybrid-A* algorithm is designed in the planning layer, incorporating multi-dimensional cost functions to satisfy motion constraints, an adaptive expansion radius to enhance search-tree flexibility, and pruning and phase-window activation strategies to further improve search efficiency. An improved Hybrid-A* algorithm is designed in the planning layer, incorporating multi-dimensional cost functions to satisfy motion constraints, an adaptive expansion radius to enhance search-tree flexibility, and pruning and phase-window activation strategies to further improve search efficiency. The structure of this paper is outlined as follows: Section 2 presents a review of the related work. Section 3 details the materials and methods employed in this study. Section 4 reports the experimental results and corresponding analysis. Section 5 summarizes the key contributions of this work and outlines promising directions for future research. 2. Related Work Path planning is fundamentally a combinatorial optimization problem that seeks an optimal solution from all feasible paths satisfying specified constraints. Early studies on path planning mainly focused on finding paths within simplified models, such as sampling-based planning algorithms and graph search methods. However, the paths produced under such simplified assumptions often overlook the practical constraints encountered by mobile robots, leading to considerable discrepancies when applied in real-world scenarios. As a result, incorporating necessary environmental, kinematic, and dynamic constraints into path planning tasks across different application domains has become an important direction of current research. Connecting the simplest coordinate sequences with straight lines often results in zigzag trajectories, causing mobile robots to frequently change motion states, leading to higher energy consumption and increased wear on motion mechanisms. In contrast, smooth path planning is better suited to the motion constraints of mobile robots. Qi et al. [15] highlight the importance of integrating heterogeneous sensors, such as wearable and environmental sensors, for improving the adaptability and accuracy of AIoT-based HAR systems, ensuring better recognition and decision-making in complex environments. One approach is to employ spline curves to smooth trajectories based on discrete waypoints generated by traditional planning algorithms. Jleilaty et al. [16] proposed a distributed control system that enhances the real-time performance and accuracy of path smoothing, addressing the challenges posed by high-curvature trajectories and collisions when waypoints are near obstacles. Elbanhawi et al. proposed a continuous path smoothing method based on B-spline curves to address continuity and maximum curvature constraints imposed by non-holonomic motion in car-like robots [17]. Blažeč et al. introduced a minimum-time optimized path-planning algorithm using low-order Bézier curves to solve motion planning problems for automated guided vehicles (AGVs) in warehouse environments, particularly considering speed and acceleration constraints [18]. Although such smoothing techniques effectively control path continuity and curvature, they rely heavily on the quality of the initial discrete waypoints. When these waypoints lie close to obstacles, the smoothing process may produce excessively high-curvature curves or lead to collisions. A second approach integrates discrete search with continuous state transitions to build continuous curve-based search trees, such as the Hybrid-A* algorithm [19] and the TargetTree-RRT* algorithm [20]. This strategy incorporates environmental and motion constraints directly into the search process. However, due to the rapid growth of nodes in the search tree, these methods are typically limited to comparatively small environments. A third approach interprets discrete waypoints from conventional algorithms as sequential local path-planning subtasks, enabling progressive exploration across multiple simplified environmental models. For example, a global dynamic path planning method combining an improved A* algorithm with the Dynamic Window Approach (DWA) was developed for optimal path generation and obstacle avoidance in complex AGV environments [21]. Similarly, Ndidiamaka et al. proposed a hybrid model that integrates Rapidly-exploring Random Tree (RRT) and DWA with a pre-trained YOLOv7 object detection model to enhance obstacle detection and avoidance within the ROS framework [22]. In addition, Wang proposed a fuzzy logic path planning method based on geometric landmarks and motion constraints, combining the Dijkstra algorithm and fuzzy logic for mobile robot navigation in complex 2D terrains [23]. However, this sequential search mechanism may still result in substantial curvature variations when switching between intermediate target points, thus limiting its effectiveness in dynamic environments. Path planning in three-dimensional environments is more suitable for complex, unstructured terrains, as it typically requires explicit 3D spatial representations, suitable constraint models for mobile robots, and comprehensive cost function formulations. Common 3D path-planning algorithms are commonly used in the field of unmanned aerial vehicles (UAVs). Afonso et al. proposed a traversability analysis and path planning technique based on mechanical effort to address autonomous navigation for unmanned ground vehicles operating in forest environments [24]. Their method extracts terrain gradient information and performs obstacle detection from 3D point cloud data, generating efficient paths avoiding unnecessary slope changes, potentially improving fuel economy, reducing equipment wear, and lowering operational risks. However, while UAVs can freely move along the vertical axis, ground mobile robots must account for both positional relationships and posture stability. Hu et al. presented an integrated path-planning system for unmanned differential-steering vehicles in off-road environments. The system utilizes 3D terrain information, obstacle potential field functions, and nonlinear optimal control strategies to efficiently generate safe and dynamically feasible trajectories [25]. Similarly, Wu et al. proposed a DEM-AIA path planning framework that integrates Digital Elevation Models with an asymmetric inclination-aware strategy based on an improved A* algorithm, enabling pitch and roll angle estimation for each newly generated triangular cell [26]. This planner also provides feasible speed references along the path, enabling efficient navigation of off-road vehicles in complex terrain. Han et al. [27] introduced an enhanced adaptive 3D path-planning algorithm (MS-W-Theta*) for mobile robots. Incorporating obstacle buffering techniques and an improved Theta* algorithm with minimum-snap trajectory smoothing, their method optimizes path planning in real-world applications, significantly enhancing path smoothness and efficiency. Mobile robots operating in unstructured environments must simultaneously address the combined effects of environmental constraints and kinematic limitations. Moreover, the curse of dimensionality in large-scale environments significantly increases the difficulty of path searching and the computational burden of planning. To effectively mitigate trajectory accumulation errors in large-scale unstructured scenarios and reduce the complexity of tracking control in high-dimensional configuration spaces, a dual-layer Hybrid A* path-planning algorithm based on phase windows is proposed. 3. Materials and Methods 3.1. Experimental Platform The experimental platform, shown in Figure 1, is an agricultural robot integrated with various sensors, such as a depth camera (D435i, Intel RealSense, Intel Corporation, Santa Clara, CA, USA), a stereo camera (ZED2, Stereolabs, San Francisco, CA, USA), a LiDAR (VLP-16, Velodyne Lidar, San Jose, CA, USA), an inertial measurement unit (SI3200G MEMS, MeiTai, Shenzhen, China), and a GNSS receiver (UM482, Unicore Communications, Beijing, China). The onboard computing and processing unit is a NUC9VXQNX workstation (Intel Corporation, Santa Clara, CA, USA; Ubuntu 20.04.6 LTS; Intel E-2286M CPU @2.4–5.0 GHz; 32 GB RAM). The ground truth for localization is provided by the NovAtel SPAN-CPT system (NovAtel Inc., Calgary, AB, Canada), an enhanced RTK+INS navigation solution. The system development is based on the Robot Operating System (ROS) Noetic and C++. The robot utilizes a four-wheel differential-drive system that enables precise control of the turning radius and navigation through varying wheel speeds. Under the operating conditions considered in this study, the minimum turning radius of the platform is approximately 0.38 m, and this value is used as the turning-radius constraint in the path-planning algorithm. 3.2. Dual-Layer Environment Model In this study, a three-fold upsampling approach is introduced to improve the resolution of the DEM, which is essential for fine-grained terrain modeling required for real-time robot operation. This method significantly enhances the accuracy of terrain features, like slope estimation and robot posture prediction, which are critical for path planning. Unlike conventional upsampling techniques that can lead to computational inefficiency or loss of accuracy, this approach strikes a balance between enhancing terrain detail and maintaining computational efficiency, ensuring real-time performance. The upsampling technique is integrated with bicubic interpolation to smooth high-frequency noise, providing a more accurate terrain representation for path planning. 3.2.1. The Digital Elevation Model (DEM) A DEM is a digital three-dimensional representation of surface elevations, commonly derived from sources like satellite imagery, LiDAR sensors, or topographic surveys. It is typically stored in a regular grid structure, where each grid point corresponds to an elevation value. DEMs provide an intuitive description of terrain variations and enable detailed analysis and computation of terrain features, such as slopes, ridges, and valleys. These models are essential for predicting and planning the posture changes experienced by mobile robots along their planned paths. The source of the DEM data in this study is field-collected point clouds using a Velodyne VLP-16 LiDAR sensor (Velodyne Lidar, San Jose, CA, USA). The native resolution of the DEM is 0.6 m per pixel, chosen to balance computational efficiency and terrain detail. Higher resolution DEMs provide more accurate representations of the terrain, but they also result in higher computational costs. A 3× bilinear interpolation upsampling is further applied to enhance the DEM resolution to 0.2 m per pixel, matching the horizontal point spacing of the VLP-16 LiDAR (≈0.17 m) and ensuring the accuracy of slope calculation for path planning. This trade-off between terrain detail and computational efficiency is a key aspect of the approach. Although higher resolution DEMs provide more accurate terrain representations, they also pose computational challenges, particularly in large-scale environments. To address this, a three-fold upsampling method enhances DEM resolution while maintaining computational feasibility for real-time robotic path planning. This balance is critical for ensuring that the path-planning algorithm remains efficient without sacrificing the accuracy of terrain features, such as slopes and ridges, which are vital for robot stability and navigation. In this study, the DEM is processed with a three-fold upsampling approach to improve the resolution for path planning. This upsampling technique increases the spatial resolution of the DEM by a factor of three, allowing for finer terrain details while maintaining the computational efficiency needed for real-time operation. The rationale behind this upsampling approach is to improve the accuracy of terrain features, such as slope estimation, which are critical for path planning, without incurring excessive computational overhead. The DEM is mathematically expressed as a two-dimensional array or matrix:(1)DEM=zij∣i=1,2,…,n;j=1,2,…,m, where zij represents the elevation value at position xi,yj. The parameters n and m represent the number of rows and columns of the matrix, respectively, defining the DEM’s spatial size. The upsampling process creates a finer resolution grid, which is then used to calculate the robot’s path more accurately. To ensure the quality of the DEM, potential noise characteristics arising from sensor errors or data acquisition limitations are considered. Simple smoothing filters, although effective at reducing noise, can blur fine terrain details, making them unsuitable for precise path planning. In contrast, this approach utilizes bicubic interpolation, which provides higher-order smoothing, effectively minimizing high-frequency noise while preserving critical terrain features. This allows for a more accurate terrain representation, especially in areas where features such as slopes and ridges significantly impact robot path planning. A three-fold upsampling approach was chosen to improve the resolution of the DEM for path planning. The rationale for this approach is twofold: first, it provides finer terrain details, such as more accurate slope estimates, which are critical for safe robot navigation in unstructured environments; second, this upsampling method enhances DEM resolution without significantly increasing computational demands. The three-fold increase in resolution strikes a balance between improving terrain representation and maintaining the computational efficiency required for real-time path planning. This approach ensures that the path-planning algorithm processes the enhanced DEM data efficiently while capturing key terrain features. 3.2.2. Construction of the Dual-Layer Environment Model In three-dimensional path planning, DEMs are typically constructed such that each grid cell is larger than the robot to ensure smooth navigation. However, this design does not account for changes in the robot’s posture during movement. To address the challenge of posture estimation for mobile robots operating in unstructured terrains, this paper proposes a two-layer environmental model. The model supports the generation of guiding points in the guide layer and the execution of path searching in the planning layer. The guide layer rapidly generates guiding points using the original DEM and establishes phase windows for subsequent planning. In bicubic interpolation, the value of each data point is influenced not only by its immediate neighboring points but also by the rate of change of adjacent samples, providing a more accurate representation of underlying trends. Compared to linear interpolation, which only considers straight-line relationships, bicubic interpolation captures more complex surface variations. This makes it particularly suitable for refining elevation maps in unstructured terrains, effectively eliminating jagged boundaries while producing smoother and more natural surfaces. Therefore, the proposed algorithm employs bicubic interpolation to construct high-resolution elevation maps for accurate posture estimation of mobile robots in the planning layer. The bicubic interpolation function at the coordinates (x,y) is given in Equation (2).(2)z(x,y)=∑i=N3∑j=N3z(xi,yj)×W(x−xi)×W(y−yj) Here, z(x,y) represents the elevation value at the point (x,y). The weight functions W(x−xi) and W(y−yj) are defined using the cubic convolution kernel proposed in [28], which achieves a balance between edge sharpness and surface smoothness through piecewise polynomial constraints. Figure 2 illustrates the elevation maps before and after resolution enhancement, where subfigure (a) shows the original low-resolution elevation map, and subfigure (b) presents the high-resolution map obtained through bicubic interpolation. The reconstructed elevation map improves the spatial resolution by a factor of three compared to the original. 3.3. Guideline Layer Establishment Method The guiding layer is responsible for generating waypoints and defining phase windows. Its primary purpose is to provide expansion directions for the search tree in the planning layer and reduce the effective search space, thereby enhancing the algorithm’s overall performance. The guiding layer workflow is as follows. First, the 3D A* algorithm is employed to obtain a sequence of guiding waypoints from the elevation map. Next, phase windows are established based on these waypoints. Finally, the departure cost within each phase window is assigned. 3.3.1. Searching for Guide Waypoints The A* algorithm, due to its optimality and computational efficiency, has become one of the most widely used graph-based path planning methods. In this study, the traditional A* algorithm is extended to a 3D multi-cost environment to obtain guiding landmarks in the guidance-layer map. The multiple cost components include distance, slope, and unknown threat. The distance cost represents the total travel distance of the mobile robot, the slope cost reflects the effect of pitch or roll during motion, and the unknown threat cost accounts for penalties incurred when traversing regions with undetectable hazards. The cost function of the improved 3D A* algorithm is expressed as follows:(3)f^(n)=g(n)+h^(n)g(n)=(ωthr+ωp+1)d+g(n−1)ωp=Δz(xn−xn−1)2+(yn−yn−1)2+(zn−zn−1)2 Here, ((xn,yn,zn)) denotes the current search position with its elevation, while ((xn−1,yn−1,zn−1)) represents the position and elevation from the previous step. The term ωp is the slope factor weight. The values g(n) and g(n−1) represent the approach costs at the current and previous search locations, respectively. The term h^(n) represents the estimated departure cost, calculated using the two-dimensional Euclidean distance between the current point and the goal point. Finally, f^(n) represents the estimated total cost at the current search location. The threat weight (ωthr) represents areas that may affect the robot’s operation, such as ponds, swamps, fires, and other hazardous regions. This weight can be adjusted based on the operating environment characteristics. In this algorithm, the threat weight is determined by the spatial distribution of multiple environmental hazards. Each hazard source located at position pti affects the threat level at point p using a spatial decay function. The cumulative threat weight is expressed as:(4)ωthr(p)=∑i=1Nωmax,i·exp−∥p−pti∥22σi2 where ωmax,i is the maximum threat weight for the i-th hazard, and σi controls its spatial influence. Guide markers are extracted at points where the direction changes along the path. Markers are classified into two categories: sparse for long, relatively straight segments, and dense for curved segments. The marker type is determined based on the distance between markers. Considering the motion characteristics of mobile robots, if the distance between markers is less than twice the minimum turning radius, they are labeled as dense; otherwise, they are labeled as sparse. 3.3.2. Phase Windows This section introduces a novel phase-window approach that combines guiding markers with Dubins curves to generate discretized, non-standard window shapes, referred to as phase windows, while similar approaches have been used in previous studies, such as coarse-to-fine path planning using fixed windows or Dubins-based regions for smoothing, this method offers dynamic adaptation by integrating real-time adjustments of window size and orientation based on environmental feedback and the robot’s kinematic constraints. In traditional methods, fixed phase windows often lead to inefficient or infeasible paths in complex or dynamic environments due to rigidity in window placement and size. On the other hand, this approach introduces sparse guiding markers, strategically placed to adjust the window configuration dynamically, ensuring more flexible and real-time path refinement. The dynamic adaptation of the phase windows enables the robot to respond to changing terrain and obstacles, providing a more robust solution to path planning in unstructured environments. Furthermore, the use of the Dubins curve to define feasible regions provides an efficient way to incorporate the minimum turning radius constraint into phase window construction. This allows the robot to accurately model its motion capabilities while avoiding excessive path curvature or infeasibilities, often seen in traditional fixed-window methods. Sparse guiding markers are distributed at relatively large intervals with minimal variation in direction. This work proposes a novel phase-window construction method based on the Dubins curve [29], while the Dubins curve is commonly used to generate feasible paths for mobile robots by connecting two points with a specified curvature, this method introduces dynamic adjustment of the guiding markers based on real-time feedback and robot posture. This enables the phase windows to adapt to terrain variations and robot constraints, improving path accuracy and robustness in unstructured environments. A Dubins curve produces a multi-segment trajectory connecting two coordinate points, given a specified curvature. Thus, it defines a feasible path for the robot to follow the guiding markers while maintaining the minimum turning radius. The region enclosed by the polyline formed by the Dubins curve and markers can be regarded as the search space that satisfies the robot’s turning-radius constraints. Mathematically, the boundary of the phase window, denoted as W, is expressed as:(5)W=(x,y)∣(x,y)∈CDubins∪CMarkers, where CDubins is the region defined by the Dubins curve, and CMarkers represents the area enclosed by the guiding markers. The robot is constrained to move within this region, ensuring that its motion adheres to the minimum turning radius constraint. The procedure for constructing dynamic phase windows with sparse guiding markers is summarized in Algorithm 1. Each phase window is generated sequentially, with local starting points assigned based on the order of the sparse guiding markers. The subsequent guiding markers are used as endpoints for generating the phase windows via Dubins curves. In traditional methods, phase windows are fixed and do not adapt to environmental changes, resulting in frequent window switching. By contrast, this approach introduces dynamic window resizing based on terrain features, ensuring smooth transitions and reducing computational overhead. Algorithm 1 Procedure for generating phase window search areas guided by sparse markers. 1:Input: Array of key points key_points 2:Output: Recorded search areas on the map 3:num_points←length(key_points) 4:order_of_points←getOrder(key_points) 5:for i←1 to num_points do 6: local_start_point ← key_points[i] 7: for j←i+1 to num_points do 8: end_point ← key_points[j] 9: dubins_curve←createDubinsCurve(local_start_point,end_point)10: search_area←encloseArea(dubins_curve,original_guide_path)11: Record end_point as the guide point12: guide_point←end_point13: updateGuidePoint(guide_point)14: if terminationConditionMet(search_area,dubins_curve) then15: recordSearchArea(search_area)16: break17: end if18: end for19:end for 1:Input: Array of key points key_points Input: Array of key points key_points 2:Output: Recorded search areas on the map Output: Recorded search areas on the map 3:num_points←length(key_points) num_points←length(key_points) 4:order_of_points←getOrder(key_points) order_of_points←getOrder(key_points) 5:for i←1 to num_points do for i←1 to num_points do 6: local_start_point ← key_points[i] local_start_point ← key_points[i] 7: for j←i+1 to num_points do for j←i+1 to num_points do 8: end_point ← key_points[j] end_point ← key_points[j] 9: dubins_curve←createDubinsCurve(local_start_point,end_point) dubins_curve←createDubinsCurve(local_start_point,end_point) 10: search_area←encloseArea(dubins_curve,original_guide_path) search_area←encloseArea(dubins_curve,original_guide_path) 11: Record end_point as the guide point Record end_point as the guide point 12: guide_point←end_point guide_point←end_point 13: updateGuidePoint(guide_point) updateGuidePoint(guide_point) 14: if terminationConditionMet(search_area,dubins_curve) then if terminationConditionMet(search_area,dubins_curve) then 15: recordSearchArea(search_area) recordSearchArea(search_area) 16: break break 17: end if end if 18: end for end for 19:end for end for The switching condition for the phase window is determined based on the following criteria: 1. Slope Variation: If the slope variation Δθ between successive guiding markers exceeds a threshold θmax, the phase window switches. This condition can be expressed as:(6)Δθ=θcurrent−θprevious>θmax. 2. Environmental Threats: If an obstacle or environmental threat Othreat is detected within a critical distance dthreat, the phase window will switch. This condition can be expressed as:(7)dthreatWmax, where Wsize is the current size of the phase window. The construction of a phase window terminates when any of the following conditions are met: 1. Significant Slope Variation: If the slope variation between the current and previous guiding markers exceeds a threshold Δθmax, the window construction is terminated. The condition can be expressed as:(9)Δθ>Δθmax. 2. Environmental Threats: If an obstacle or threat is detected within a critical distance dcritical, the phase window construction terminates to avoid collisions:(10)dthreatWmax. Once any of these conditions are met, the system proceeds to generate the next phase window, ensuring that the path remains feasible and efficient (Figure 3). 3.3.3. Heuristic Cost in Phase Windows The heuristic cost guides the path-planning algorithm in the planning layer. The guiding markers constructed here store the accumulated cost from the starting point to each marker. The heuristic cost at each guiding marker is obtained by computing the difference between the endpoint cost and the marker cost. The heuristic cost values for other positions in a phase window are then computed based on these guiding markers. (a)Heuristic cost in phase windows guided by sparse markers Heuristic cost in phase windows guided by sparse markers The heuristic cost h(xd,yd) at a coordinate (xd,yd) within a phase window constructed using sparse guiding markers is calculated using Equation (12).(12)h^(xd,yd)=h^(xls,yls)DISgpDISls+DISgp+h^(xgp,ygp)DISlsDISls+DISgpDISgp=(xgp−xd)2+(ygp−yd)2DISls=(xd−xls)2+(yd−yls)2 gx_(n)=gx_(n−1)+∑i=2m1ωgrad+ϵ+1ωbump+ϵ+1ωthr+ϵgdis Here, (xls,yls) denotes the starting marker of the phase window, and (xgp,ygp) denotes the ending guiding marker. When phase windows overlap, the smaller heuristic cost is recorded. This ensures the heuristic cost consistently decreases toward the global endpoint. (b)Heuristic cost in phase windows guided by dense markers Heuristic cost in phase windows guided by dense markers The phase window guided by dense markers contains three points (Pi−1,Pi,Pi+1). By projecting Pi−1 and Pi perpendicularly onto vectors Pi−1Pi→ and PiPi+1→, the phase window is divided into three subregions (as illustrated in Figure 4). Coordinates falling within the invalid region are excluded from the heuristic cost computation. For the remaining two valid regions, the heuristic cost is computed using Equation (12). In the first region, Pi is the starting marker and Pi+1 the ending marker. In the second region, Pi−1 is the starting marker and Pi the ending marker. (c)Heuristic cost in special phase windows Heuristic cost in special phase windows At the starting guiding marker, besides the phase window region defined by sparse markers, the heuristic cost in the supplementary square region is the sum of the start point’s heuristic cost and the Euclidean distance from the start point to the coordinate. At the ending guiding marker, except where heuristic costs are already assigned in other phase windows, the heuristic cost in the supplementary square region is defined as the Euclidean distance from the coordinate to the endpoint. Thus, the search direction in the planning layer always guides the robot toward the global endpoint, as summarized in Algorithm 2. Algorithm 2 Heuristic cost calculation in phase windows. 1:Input: Guide markers and coordinate (xd,yd) 2:Output: Heuristic cost h(xd,yd) 3:if phase window guided by sparse markers then 4: Compute DISgp and DISls 5: Compute h(xd,yd) using the weighted-average formula (Equation (12)) 6: if overlap exists between phase windows then 7: Record the minimum value of h(xd,yd) 8: end if 9:end if10:if phase window guided by dense markers then11: Partition the phase window into subregions12: if (xd,yd) lies in a valid subregion then13: Compute h(xd,yd) based on the corresponding guide markers14: end if15:end if16:if special phase window at the start or end marker then17: if (xd,yd) is at the start-marker region then18: h(xd,yd)=hstart+EuclideanDistance19: else if (xd,yd) is at the end-marker region then20: h(xd,yd)=EuclideanDistancetoendpoint21: end if22:end if23:Return h(xd,yd) 1:Input: Guide markers and coordinate (xd,yd) Input: Guide markers and coordinate (xd,yd) 2:Output: Heuristic cost h(xd,yd) Output: Heuristic cost h(xd,yd) 3:if phase window guided by sparse markers then if phase window guided by sparse markers then 4: Compute DISgp and DISls Compute DISgp and DISls 5: Compute h(xd,yd) using the weighted-average formula (Equation (12)) Compute h(xd,yd) using the weighted-average formula (Equation (12)) 6: if overlap exists between phase windows then if overlap exists between phase windows then 7: Record the minimum value of h(xd,yd) Record the minimum value of h(xd,yd) 8: end if end if 9:end if end if 10:if phase window guided by dense markers then if phase window guided by dense markers then 11: Partition the phase window into subregions Partition the phase window into subregions 12: if (xd,yd) lies in a valid subregion then if (xd,yd) lies in a valid subregion then 13: Compute h(xd,yd) based on the corresponding guide markers Compute h(xd,yd) based on the corresponding guide markers 14: end if end if 15:end if end if 16:if special phase window at the start or end marker then if special phase window at the start or end marker then 17: if (xd,yd) is at the start-marker region then if (xd,yd) is at the start-marker region then 18: h(xd,yd)=hstart+EuclideanDistance h(xd,yd)=hstart+EuclideanDistance 19: else if (xd,yd) is at the end-marker region then else if (xd,yd) is at the end-marker region then 20: h(xd,yd)=EuclideanDistancetoendpoint h(xd,yd)=EuclideanDistancetoendpoint 21: end if end if 22:end if end if 23:Return h(xd,yd) Return h(xd,yd) 3.4. Planning Layer Establishment Method The planning layer integrates an enhanced Hybrid-A* algorithm that performs path planning within the phase windows generated by the guidance layer, ensuring compliance with mobile robot motion constraints. First, the algorithm introduces an improved cost function adaptable to complex, unstructured environments. Second, it incorporates an adaptive turning-radius selection mechanism to prevent unnecessary turns or loops caused by a fixed turning radius. Finally, a pruning strategy is employed to accelerate the path search and reduce computational overhead. 3.4.1. Cost Function The cost function of the improved Hybrid-A* algorithm is given in Equation (13).(13)f^(x_(n))=g(x_(n))+h^(x_(n))g(x_(n))=g(x_(n−1))+∑i=2m1+ωgradωgrad,max+ϵ+ωbumpωbump,max+ϵ+ωthrωthr,max+ϵgdis In this formulation, the heuristic cost h^(xn) represents the estimated cost to reach the endpoint, whereas the approach cost g(xn) denotes the accumulated cost from start to current position. The term gdis denotes the distance cost for each subdivided segment of the expansion tree, and m is the number of subdivisions per segment. The gradient weight ωgrad represents the inclination of the robot on the terrain; the bumpiness weight ωbump quantifies the robot’s vibration intensity; and the threat weight ωthr reflects the potential risk associated with unknown environmental hazards. To ensure the stability and generalizability of the cost function, each weight term is normalized by dividing it by its maximum value (e.g., ωgrad,max) plus a small constant ϵ to prevent division by zero. Normalization standardizes the range of values for ωgrad, ωbump, and ωthr, ensuring that no single term dominates the total cost due to differences in their physical units or magnitudes. (a)Distance cost (gdis) Distance cost (gdis) The distance cost represents the actual traveling distance and is computed as the three-dimensional Euclidean distance between two adjacent points:(14)gdis=(xi−xi−1)2+(yi−yi−1)2+(zi−zi−1)2 (b)Gradient weight (ωgrad) Gradient weight (ωgrad) The gradient weight is calculated based on the robot’s physical model and local terrain conditions, allowing the algorithm to evaluate the robot’s inclination at each position. Integrating the gradient weight into the cost function helps prevent slipping, tipping, and energy wastage on steep terrain.(15)ωgrad=tan(max(θf−b,θl−r,θfl−br,θfr−bl))θf−b=tan−1zfl+zfr2−zbl+zbr2(xfl−xbl)2+(yfl−ybl)2θl−r=tan−1zfl+zbl2−zfr+zbr2(xfl−xfr)2+(yfl−yfr)2θfl−br=tan−1zfl−zbr(xfl−xbr)2+(yfl−ybr)2θfr−bl=tan−1zfr−zbl(xfr−xbl)2+(yfr−ybl)2 The spatial coordinates of the four wheels of the mobile robot are defined as:Pleft-front-wheelPleft-rear-wheelPright-front-wheelPright-rear-wheel=xflyflzflxblyblzblxfryfrzfrxbrybrzbr (c)Bumpiness weight ωbump Bumpiness weight ωbump The bumpiness weight evaluates robot posture changes along the path segment. Incorporating this weight helps reduce mechanical wear and localization errors due to excessive vibration. To ensure stability and minimize abrupt variations, a smoothing strategy for the gradient differences is introduced. The bumpiness weight ωbump,i is computed from the smoothed gradient differences between consecutive nodes, so larger values correspond to stronger posture fluctuations along the path.(16)ωbump=NnowNmax,Nnow≤Nmaxα·Nnow,Nnow>NmaxNnow=ωgrad,i−ωgrad,i−1·wsmoothNmax=f(robot’smaxslope,terraincharacteristics,orexperimentaldata) Here, ωgrad,i denotes the gradient weight at the i-th node of the expansion tree, and Nmax represents the maximum allowable change in gradient for safe robot operation. When the local gradient change exceeds Nmax, the penalty associated with the bumpiness term is saturated at its upper bound, so that excessively large variations do not lead to unbounded costs. The smoothing factor wsmooth helps reduce abrupt gradient variations by applying a weighted average of the gradient differences. This ensures that the bumpiness term remains stable and does not destabilize the search process. 3.4.2. Adaptive Turning Radius This section introduces two methods for selecting an adaptive turning radius. Unlike traditional fixed-radius search strategies, these methods allow more flexible determination of the search tree’s expansion direction based on guidance layer information. (a)Guide point radius Guide point radius First, the method is called the guide point radius. A circular arc is constructed from the current search position (xd,yd) and heading direction (θd) toward the guide point (xgp,ygp). This arc is called the guide point arc, and its radius is termed the guide point radius. If the guide point radius exceeds the minimum turning radius of the mobile robot, the search tree expands in that direction (Figure 5). Equation (17) calculates the guide point radius.(17)rgp=(xgp−xd)2+(ygp−yd)22sintan−1ygp−ydxgp−xd−θd+ϵ To prevent numerical instability when the heading aligns closely with the target direction, a small constant ϵ is added to the denominator. This ensures that the denominator never approaches zero, thus stabilizing the guide point radius computation. (b)Guide direction radius Guide direction radius The direction from the local start point to its guide point is defined as the guide direction. The angle between the guide direction and the robot’s current heading is termed the adjustment angle. From the current position, an arc can be drawn using the adjustment angle as the central angle and the search step as the arc length. This arc is referred to as the guide direction arc, and its radius is the guide direction radius. If this radius exceeds the minimum turning radius of the mobile robot, the search tree is expanded accordingly (Figure 5). Equation (18) computes the guide direction radius.(18)rgd=Lθ−ψ⌢tan(ψ)=ygp−ylsxgp−xls Here, θ and ψ denote the current search direction and the guidance direction, respectively. L represents the search step size. (xgp,ygp) is the coordinate of the guide point, and (xls,yls) is the coordinate of the local start point. 3.4.3. Pruning Method The pruning method is designed to address two issues during path search: (1) the generation of unnecessary search branches that deviate from the expected search direction, and (2) the repeated generation of search nodes in locally dense regions, leading to redundant nodes with similar positions and headings. To mitigate these problems, the following pruning strategies are adopted:Phase window constraint: Nodes outside the active phase window are pruned to restrict the search to dynamically defined feasible regions.Directional consistency: Nodes forming an obtuse angle with direction ψ are removed to prevent reverse or divergent expansion of the search tree.Redundancy suppression: New nodes are discarded if their distance from an existing node is smaller than the robot’s width and heading difference is Δθ<7.5∘, reducing local redundancy. Phase window constraint: Nodes outside the active phase window are pruned to restrict the search to dynamically defined feasible regions. Phase window constraint: Nodes outside the active phase window are pruned to restrict the search to dynamically defined feasible regions. Directional consistency: Nodes forming an obtuse angle with direction ψ are removed to prevent reverse or divergent expansion of the search tree. Directional consistency: Nodes forming an obtuse angle with direction ψ are removed to prevent reverse or divergent expansion of the search tree. Redundancy suppression: New nodes are discarded if their distance from an existing node is smaller than the robot’s width and heading difference is Δθ<7.5∘, reducing local redundancy. Redundancy suppression: New nodes are discarded if their distance from an existing node is smaller than the robot’s width and heading difference is Δθ<7.5∘, reducing local redundancy. 3.4.4. Activation Method for Phase Windows Phase windows are a sequence of guided search regions generated by the guidance layer, each annotated with heuristic costs. In the planning layer, the expansion tree grows exponentially, which can severely reduce search efficiency in large-scale environments. To alleviate this, this section proposes an activation mechanism based on phase window activity levels, aiming to reduce candidate nodes during the search. Assume that the global map contains n phase windows. Let Wini be the phase window containing the current node. Based on its index position, the state of each window is defined as follows:(19)Windowsj∈Active,i−10≤j≤nSemi-active,i−20≤j