Multi-amr collaborative carrying obstacle avoidance method in dynamic scene

Abstract

The application discloses a multi-AMR cooperative carrying obstacle avoidance method, which comprises the following steps: after receiving a task, an AMR is allocated and an initial priority is set; a dynamic cost map is constructed by fusing a static map and a history passing data exponential decay heat matrix; a heat penalty is added in path searching to avoid congestion; dynamic high-speed objects are perceived by a laser radar and a depth camera during driving; after visual classification, a motion prediction model is called to generate a predicted trajectory; based on the space-time conflict detection of the path of the AMR and the predicted trajectory, resource blocks are divided to obtain an entering sequence and a time window when other AMRs are involved; an occupancy directed graph is constructed to detect a deadlock; only when a general conflict exists, autonomous re-planning is performed to bypass; when there is a risk of deadlock, an edge node is triggered to arbitrate cooperative avoidance according to the priority and the waiting time; when the arbitration is not received within a time limit, the arbitration is downgraded to autonomous obstacle avoidance according to the local priority; after reaching, docking is performed; and the stability is confirmed and reported by using a limit switch and an IMU.

Description

Technical Field

[0001] This invention belongs to the field of multi-machine collaborative scheduling and control technology for mobile robots, specifically relating to a multi-AMR collaborative transport and obstacle avoidance method in dynamic scenarios. Background Technology

[0002] When performing collaborative transport tasks, existing AMRs mostly use static maps for global path planning. When encountering dynamic obstacles, they rely solely on single-machine reactive obstacle avoidance. When multiple AMRs meet at intersections or narrow passages, there is a lack of conflict prediction and deadlock detection mechanisms for the future paths of each AMR. Furthermore, it is impossible to adopt differentiated avoidance strategies based on the severity of the conflict, which easily leads to deadlock phenomena of multiple machines waiting in a loop. In addition, the communication and computing resource overhead is large, and the system's passage efficiency is low. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-AMR cooperative transport obstacle avoidance method in dynamic scenarios to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a multi-AMR cooperative transport and obstacle avoidance method in dynamic scenarios, comprising the following steps:

[0005] Receive handling tasks, assign tasks to each AMR and set initial priorities;

[0006] For each AMR, a dynamic cost map is constructed based on a static map overlaid with a heat matrix generated from historical traffic data and maintained using exponential decay. The cost function of the path search algorithm introduces a penalty term into the heat matrix to plan a global reference path that can proactively avoid historical congestion areas.

[0007] Each AMR travels along the global reference path and simultaneously perceives the environment in real time through onboard LiDAR and depth camera. It detects objects with a speed higher than a preset threshold and identifies obstacle types through a visual classification model. The types include at least pedestrians, other AMRs, and temporary materials. Based on different obstacle types, it calls the corresponding motion prediction model to generate predicted trajectories.

[0008] Spatiotemporal conflict detection is performed based on the planned path and predicted trajectory of obstacles of this AMR. When the conflict involves other AMRs, the conflict area is divided into multiple resource blocks. The planning order and time window of each related AMR entering each resource block are obtained. A directed graph with AMRs as nodes and resource occupation waiting relationship as edges is constructed. The presence of deadlock risk is determined by detecting loops on the directed graph.

[0009] If only general conflicts that are not deadlocked exist, the AMR will autonomously perform local path replanning and decelerate to detour. The local path replanning is guided by the global reference path.

[0010] If a deadlock risk exists, an edge-to-edge collaborative obstacle avoidance strategy is triggered. The edge node arbitrates based on the priority and waiting time of each AMR and generates a collaborative instruction. The relevant AMR executes the collaborative instruction to resolve the deadlock. If an AMR does not receive an arbitration instruction within a preset timeout after sending a deadlock report to the edge node, it automatically degrades to an autonomous obstacle avoidance mode that determines the yielding order based on the priority of its local cache.

[0011] Once the AMR reaches the target workstation, it performs the handling and docking action, and after confirming the stability of the handling through limit switches and inertial measurement units, it reports the task as completed.

[0012] Preferably, the maintenance method of the thermal matrix includes:

[0013] Each grid records the cumulative number of times the AMR passes through within a preset time period, and performs exponential decay processing according to a preset period. The new thermal value is the old thermal value multiplied by the decay coefficient of 0.5 and then added to the count added in the current statistical period.

[0014] The cost function is f(n)=g(n)+h(n)+α×HeatMap[n]×dist_factor, where α is the weight of thermal influence and dist_factor is the attenuation factor that varies with the distance from the node to the target point, so as to avoid excessive detours on road segments near the end of the path.

[0015] Preferably, the step of generating a predicted trajectory by calling the corresponding motion prediction model according to different obstacle types includes:

[0016] For pedestrians, a constant velocity model combined with social force correction is used for trajectory prediction. The social force correction is used to simulate the pedestrian's tendency to keep a distance from other targets.

[0017] For other AMRs, their current target path information is obtained through V2X communication, and their linear velocity and angular velocity are combined to perform trajectory extrapolation using a motion model with constant curvature and constant acceleration.

[0018] Temporary materials are treated as stationary obstacles with fixed positions.

[0019] Preferably, the construction of a directed graph with AMRs as nodes and resource occupancy / wait relationships as edges, and the determination of deadlock risk by detecting loops in the directed graph, includes:

[0020] If a resource block planned to be occupied by one AMR is reserved by another AMR and the occupation time periods of the two overlap, then add an edge from the first AMR to the other AMR.

[0021] Run a depth-first search algorithm on the constructed directed graph to detect loops. If a loop exists, it is considered a deadlock risk, and the AMRs contained in the loop are identified as participants in the cooperative avoidance.

[0022] If there are no cycles but there are edges, it is considered a general conflict that is not deadlocked.

[0023] Preferably, the step of autonomously performing local path replanning and deceleration by the AMR, wherein the local path replanning is guided by the global reference path, includes:

[0024] Candidate trajectories are generated by sampling in the velocity space using the dynamic window method. The progress component toward the global path in the evaluation function is forced to be guided by the global reference path planned based on the dynamic cost map, so that the local detour trajectory tends to avoid historical congestion areas while avoiding obstacles.

[0025] Select the optimal trajectory and reduce the AMR's speed limit from the current value to a preset safe value until it safely passes through the conflict area and then resumes the original speed.

[0026] Preferably, the step of arbitrating and generating coordination instructions by the edge nodes based on the priority and waiting time of each AMR includes:

[0027] Compare the task priorities of each AMR involved in the deadlock, and the task with the lowest priority value will have priority to pass.

[0028] If there are AMRs with the same priority, then the accumulated waiting time of each AMR in the conflict area is further compared, and the AMR with the longer waiting time is given priority to pass.

[0029] For AMRs that need to yield, an instruction is generated to pause yielding or back up to the avoidance point based on their current position and whether there are other AMRs behind them.

[0030] Preferably, the arbitration and generation of collaborative instructions further includes:

[0031] If there is an AMR with a task priority lower than the preset threshold and the conflict area space is insufficient to accommodate all AMRs that need to be paused, then a global path replanning instruction is directly issued to the low-priority AMR, so that it actively avoids the current conflict area.

[0032] Preferably, the automatic degradation is an autonomous obstacle avoidance mode that determines the yielding order based on the priority of the local cache, including:

[0033] If each AMR involved in deadlock does not receive an arbitration instruction from the edge node within a preset timeout period, it will read the priority information from its local cache.

[0034] Those with lower priority values ​​pass through the conflict zone first, while those with higher priority values ​​either slow down to give way or pause and wait.

[0035] If the local caches have the same priority, the passage order is determined by sorting the device identifiers in the AMR.

[0036] Preferably, the visual classification model is a lightweight target detection and instance segmentation model deployed on an AMR vehicle-mounted embedded computing unit. The model receives the feature fusion results of the point clusters after LiDAR point cloud clustering and the RGB images of the depth camera as input, and outputs the type and contour information of the obstacle.

[0037] Preferably, the step of reporting task completion after confirming stable handling via limit switches and inertial measurement units includes:

[0038] Confirm that the lifting mechanism has reached the target position by using the limit switch;

[0039] The roll or pitch angle deviation of the current attitude is detected by the inertial measurement unit and does not exceed a preset threshold.

[0040] Once the above conditions are met, the loading status is locked, and then the task is reported as completed to the warehouse management system or manufacturing execution system.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] The system receives transport tasks, assigns tasks to each AMR, and sets initial priorities. For each AMR, a heatmap generated from historical traffic data and maintained using exponential decay is overlaid on a static map to construct a dynamic cost map. The cost function of the path search algorithm incorporates a penalty term from the heatmap to plan a global reference path that can proactively avoid historical congestion areas. Each AMR travels along the global reference path, simultaneously sensing the environment in real time using onboard LiDAR and depth cameras. Objects with movement speeds exceeding a preset threshold are detected, and obstacle types are identified using a visual classification model. These types include at least pedestrians, other AMRs, and temporary materials. Based on different obstacle types, corresponding motion prediction models are invoked to generate predicted trajectories. Spatiotemporal conflict detection is performed based on the planned path of this AMR and the predicted trajectories of obstacles. When a conflict involves other AMRs, the conflict area is divided into multiple resource blocks. The planning order and time window for each relevant AMR entering each resource block are obtained, and a directed graph is constructed with AMRs as nodes and resource occupancy waiting relationships as edges. The presence of deadlock risk is determined by detecting loops in the directed graph. If only general conflicts that are not deadlock-related exist, the AMR will autonomously perform local path replanning and decelerate to avoid the deadlock, with the global reference path serving as the guiding reference. If there is a risk of deadlock, an edge-to-edge cooperative obstacle avoidance strategy will be triggered. The edge node will arbitrate based on the priority and waiting time of each AMR and generate a cooperative instruction. The relevant AMRs will execute the cooperative instruction to resolve the deadlock. If an AMR does not receive an arbitration instruction within a preset timeout after sending a deadlock report to the edge node, it will automatically downgrade to an autonomous obstacle avoidance mode where it determines the yielding order based on the priority of its local cache. After the AMR reaches the target workstation, it will perform a transport and docking operation, and after confirming stable transport through limit switches and an inertial measurement unit, it will report the task as completed. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the present invention.

[0044] Figure 2 This is an interactive diagram of thermal matrix maintenance and global path planning in this invention.

[0045] Figure 3 This is a flowchart illustrating the obstacle classification and trajectory prediction process of the present invention.

[0046] Figure 4 This is a flowchart illustrating the spatiotemporal conflict detection and deadlock determination process of the present invention.

[0047] Figure 5 This is a schematic diagram of the autonomous local path replanning process under non-deadlock conflict of the present invention.

[0048] Figure 6 This is a sequence diagram of edge-to-edge collaborative avoidance under deadlock risk in this invention.

[0049] Figure 7 This is a sequence diagram of the edge arbitration timeout degradation fault tolerance of the present invention.

[0050] Figure 8 This is a sequence diagram of the transportation, docking, and completion confirmation of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1:

[0053] A multi-AMR cooperative transport obstacle avoidance method in dynamic scenarios includes the following steps:

[0054] Receive handling tasks, assign tasks to each AMR and set initial priorities;

[0055] For each AMR, a dynamic cost map is constructed based on a static map overlaid with a heat matrix generated from historical traffic data and maintained using exponential decay. The cost function of the path search algorithm introduces a penalty term into the heat matrix to plan a global reference path that can proactively avoid historical congestion areas.

[0056] Each AMR travels along a global reference path and simultaneously perceives the environment in real time through onboard LiDAR and depth cameras. It detects objects with a speed exceeding a preset threshold and identifies obstacle types through a visual classification model. The types include at least pedestrians, other AMRs, and temporary materials. Based on different obstacle types, it calls the corresponding motion prediction model to generate predicted trajectories.

[0057] Spatiotemporal conflict detection is performed based on the planned path and predicted trajectory of obstacles of this AMR. When the conflict involves other AMRs, the conflict area is divided into multiple resource blocks. The planning order and time window of each related AMR entering each resource block are obtained. A directed graph with AMRs as nodes and resource occupation waiting relationship as edges is constructed. The presence of deadlock risk is determined by detecting loops on the directed graph.

[0058] If only general conflicts that are not deadlocked exist, the AMR will autonomously perform local path replanning and decelerate to detour. The local path replanning is guided by the global reference path.

[0059] If there is a deadlock risk, the edge-to-edge collaborative obstacle avoidance strategy is triggered. The edge node arbitrates based on the priority and waiting time of each AMR and generates a collaborative instruction. The relevant AMR executes the collaborative instruction to resolve the deadlock. If the AMR does not receive the arbitration instruction within the preset timeout after sending the deadlock report to the edge node, it automatically degrades to the autonomous obstacle avoidance mode, which determines the yielding order based on the priority of the local cache.

[0060] Once the AMR reaches the target workstation, it performs the handling and docking operation. After confirming stable handling via limit switches and the inertial measurement unit, it reports task completion. The maintenance methods for the thermal matrix include:

[0061] Each grid records the cumulative number of times the AMR passes through within a preset time period, and performs exponential decay processing according to a preset period. The new thermal value is the old thermal value multiplied by the decay coefficient of 0.5 and then added to the count added in the current statistical period.

[0062] The cost function is f(n) = g(n) + h(n) + α × HeatMap[n] × dist_factor, where α is the weight of the thermal effect, and dist_factor is a decay factor that varies with the distance from the node to the target point to avoid excessive detours on road segments near the path endpoint. Based on different obstacle types, the corresponding motion prediction model is called to generate the predicted trajectory, including:

[0063] For pedestrians, a constant velocity model combined with social force correction is used for trajectory prediction. Social force correction is used to simulate the pedestrian's tendency to keep a distance from other targets.

[0064] For other AMRs, their current target path information is obtained through V2X communication, and their linear velocity and angular velocity are combined to perform trajectory extrapolation using a motion model with constant curvature and constant acceleration.

[0065] For temporary materials, they are treated as stationary obstacles with fixed positions. A directed graph is constructed with AMRs as nodes and resource occupancy and waiting relationships as edges. Deadlock risks are assessed by detecting loops in the directed graph, including:

[0066] If a resource block planned to be occupied by one AMR is reserved by another AMR and the occupation time periods of the two overlap, then add an edge from the first AMR to the other AMR.

[0067] Run a depth-first search algorithm on the constructed directed graph to detect loops. If a loop exists, it is considered a deadlock risk, and the AMRs contained in the loop are identified as participants in the cooperative avoidance.

[0068] If there are no loops but there are edges, it is considered a general conflict that is not deadlocked. This AMR autonomously performs local path replanning and decelerates to avoid a loop. The local path replanning is guided by the global reference path and includes:

[0069] The dynamic window method is used to sample and generate candidate trajectories in the velocity space. The progress component towards the global path in the evaluation function is forced to be guided by the global reference path planned based on the dynamic cost map, so that the local detour trajectory tends to avoid historical congestion areas while avoiding obstacles.

[0070] The optimal trajectory is selected for execution, while the AMR's maximum speed is reduced from its current value to a preset safety value until it safely passes through the conflict zone, at which point the original speed is restored. Edge nodes arbitrate based on the priority of each AMR and the waiting time, and generate coordinated instructions, including:

[0071] Compare the task priorities of each AMR involved in the deadlock, and the task with the lowest priority value will have priority to pass.

[0072] If there are AMRs with the same priority, then the accumulated waiting time of each AMR in the conflict area is further compared, and the AMR with the longer waiting time is given priority to pass.

[0073] For AMRs that need to yield, based on their current position and whether there are other AMRs behind them, an instruction is generated to pause yielding or reverse to a passing point. Arbitration and generation of coordination instructions also include:

[0074] If an AMR with a task priority lower than a preset threshold exists and the conflict zone space is insufficient to accommodate all AMRs that need to be paused, a global path replanning instruction is directly issued to the low-priority AMR, causing it to actively avoid the current conflict zone. Automatic degradation to an autonomous obstacle avoidance mode that determines the yielding order based on the priority of the local cache includes:

[0075] If each AMR involved in deadlock does not receive an arbitration instruction from the edge node within a preset timeout period, it will read the priority information from its local cache.

[0076] Those with lower priority values ​​pass through the conflict zone first, while those with higher priority values ​​either slow down to give way or pause and wait.

[0077] If local cache priorities are the same, the passage order is determined by the AMR's device identifier. The visual classification model is a lightweight target detection and instance segmentation model deployed on the AMR's onboard embedded computing unit. The model receives the feature fusion results of point clusters after LiDAR point cloud clustering and RGB images from depth cameras as input, and outputs obstacle type and contour information. After confirming stable handling via limit switches and inertial measurement units, the task is reported as complete, including:

[0078] Confirm that the lifting mechanism has reached the target position by using the limit switch;

[0079] The roll or pitch angle deviation of the current attitude is detected by the inertial measurement unit and does not exceed a preset threshold.

[0080] Once the above conditions are met, the loading status is locked, and then the task is reported as completed to the warehouse management system or manufacturing execution system.

[0081] Through the above technical solution, after receiving a handling task, this invention first assigns tasks to each AMR and sets an initial priority. For each AMR, a heat map generated from historical traffic data is overlaid on a static map. The heat map is maintained using exponential decay to construct a dynamic cost map. The cost function of path search incorporates a heat penalty term, and a global reference path is planned to avoid historical congestion areas. Each AMR travels along the global reference path, while simultaneously using onboard LiDAR and depth cameras to perceive the environment in real time, detecting objects with speeds exceeding a preset threshold. A visual classification model is used to identify pedestrians, other AMRs, or temporary materials, and the corresponding motion prediction model is called to generate predicted trajectories. Spatiotemporal conflict detection is performed based on the planned path of this AMR and the predicted obstacle trajectory. If the conflict involves other AMRs, the conflict area is divided into multiple resource blocks, and the order and time window of each AMR entering each resource block are recorded. A directed graph is constructed with AMRs as nodes and resource occupancy waiting relationships as edges. The presence of deadlock risk is determined by detecting loops in the graph. If it is only a general conflict, this AMR autonomously performs local path replanning and decelerates to detour, guided by the global reference path during replanning. If a deadlock risk exists, edge-to-edge collaborative obstacle avoidance is triggered. The edge node arbitrates based on the priority and waiting time of each AMR and issues a collaborative command. The relevant AMR executes the command to resolve the deadlock. If an AMR sends a deadlock report to the edge node but does not receive an arbitration command within a preset time, it automatically degrades to an autonomous obstacle avoidance mode where it determines the yielding order based on the priority of its local cache. After the AMR reaches the target workstation, it performs transport docking, confirms stable transport through limit switches and inertial measurement units, and then reports task completion. By using on-demand scheduling, the system avoids strong real-time dependence on the central controller, reduces overall system communication and computing overhead, and prevents multiple AMRs from entering a cyclical waiting deadlock at intersections, narrow passages, etc., significantly improving traffic efficiency and task execution stability under high-concurrency operations.

[0082] Example 2:

[0083] In this embodiment, during the operation of the multi-AMR cooperative transport and obstacle avoidance method in dynamic scenarios, the scheduling system receives transport tasks, assigns tasks to each AMR, and sets initial priorities. For each AMR, the processor calls the static map in the storage unit and reads historical passage data from the historical passage data storage module. The historical passage data is generated by a heat matrix maintained by the periodically reported passage records of each AMR through an exponential decay algorithm. The exponential decay algorithm continuously reduces the weight of historical data with a time factor. The processor overlays the heat matrix onto the static map to construct a dynamic cost map. When the path search algorithm expands nodes, the cost function introduces a penalty term for the corresponding grid in the heat matrix. The value of the penalty term is positively correlated with the heat value, thereby planning a global reference path that can actively avoid historical congestion areas. The global reference path is stored in the AMR's local memory.

[0084] Each AMR travels along a global reference path. The onboard LiDAR emits laser beams and receives reflected signals, while the depth camera captures environmental depth images. Both input the perceived data into the onboard computing unit. The onboard computing unit estimates the velocity of objects in the perceived data, filtering out objects with speeds exceeding a preset threshold. A visual classification model then identifies the types of these filtered objects, with the identification results including at least pedestrians, other AMRs, and temporary materials. Based on the identified obstacle type, the onboard computing unit invokes the corresponding motion prediction model. This model generates a predicted trajectory based on the obstacle's current motion state, and the predicted trajectory includes a timestamp and a sequence of spatial coordinates.

[0085] The onboard computing unit performs spatiotemporal conflict detection between the planned path of the current AMR and the predicted trajectories of obstacles. When a conflict involves other AMRs, the onboard computing unit divides the conflict area into multiple resource blocks and obtains the planning order and time window for each related AMR to enter each resource block. The onboard computing unit constructs a directed graph with AMRs as nodes and resource occupancy waiting relationships as edges. If AMR-A needs to wait for AMR-B to release a resource block, a directed edge is generated from node A to node B. The onboard computing unit runs a loop detection algorithm on the directed graph; if a loop exists, a deadlock risk is identified.

[0086] If only general conflicts that are not deadlock-free exist, the onboard computing unit performs local path replanning, which generates a detour path based on the global reference path. Simultaneously, the onboard controller reduces the AMR's speed. If a deadlock risk exists, the AMR's communication module sends a deadlock report to the edge node. The edge node's arbitration module arbitrates based on the priority of each AMR and the waiting time, generates a coordination command, and sends it to the relevant AMR via the wireless communication module. The onboard controller of the relevant AMR executes the coordination command to resolve the deadlock. If the AMR's communication module does not receive an arbitration command within a preset timeout after sending the deadlock report to the edge node, the onboard computing unit automatically downgrades to autonomous obstacle avoidance mode, reads the cached priorities from local memory, and determines the yielding order based on the priorities.

[0087] Once the AMR reaches the target workstation, the onboard controller drives the actuator to perform the transport and docking action. The limit switch detects whether the docking parts have reached the predetermined position and outputs a switch signal. The inertial measurement unit measures the AMR's attitude angle and acceleration data. The onboard computing unit receives the limit switch signal and the inertial measurement unit data. After confirming that the transport is stable, the communication module reports the task completion to the scheduling system.

[0088] Example 3:

[0089] In this embodiment, the maintenance method of the multi-AMR cooperative transport obstacle avoidance method in dynamic scenarios is as follows: each grid records the cumulative number of times the AMR passes through within a preset time period, and performs exponential decay processing according to a preset period. The new heat value is the old heat value multiplied by the decay coefficient of 0.5 and then added to the count added in the current statistical period. The cost function adopts the form f(n)=g(n)+h(n)+α×HeatMap[n]×dist_factor, where α is the heat influence weight and dist_factor is the decay factor that varies with the distance from the node to the target point, so as to avoid excessive detours on road segments near the end of the path.

[0090] During operation, the onboard computing units of each AMR overlay a heat map generated from historical traffic data onto a static map to construct a dynamic cost map. Each grid cell of the heat map continuously maintains the cumulative number of times the AMR has passed through the area within a preset time period. The system triggers exponential decay processing according to a preset cycle, multiplying the old heat map value by a decay coefficient of 0.5 and adding the new count added in the current statistical cycle to obtain the new heat map value. When expanding nodes, the path search algorithm calculates the cost function f(n) = g(n) + h(n) + α × HeatMap[n] × dist_factor, where g(n) is the actual cost from the starting point to the current node, h(n) is the heuristically estimated cost from the current node to the target point, α is the heat map influence weight, HeatMap[n] is the heat map value corresponding to node n, and dist_factor is the decay factor that varies with the distance from the node to the target point. dist_factor approaches zero when the node is close to the target point, which reduces the impact of the heat map penalty on road segments near the end of the path and avoids unnecessary detours. The AMR travels along a planned global reference path, which actively avoids historically congested areas due to the inclusion of a thermal penalty term in the cost function. When the AMR perceives the environment in real time using its onboard LiDAR and depth camera and detects objects moving at speeds exceeding a preset threshold, a visual classification model identifies the obstacle type and calls the corresponding motion prediction model to generate a predicted trajectory. Spatiotemporal conflict detection is performed based on the AMR's planned path and the predicted trajectories of obstacles. If the conflict involves other AMRs, the conflict area is divided into multiple resource blocks. The planning order and time window for each relevant AMR entering each resource block are obtained, and a directed graph is constructed with AMRs as nodes and resource occupancy waiting relationships as edges. Cycle detection is performed on the directed graph to determine if deadlock risk exists. If only general conflicts that do not result in deadlock exist, the AMR autonomously performs local path replanning and decelerates to avoid the congestion, with the global reference path serving as the guiding benchmark. If deadlock risk exists, an edge-to-edge collaborative avoidance strategy is triggered. Edge nodes arbitrate based on the priority and waiting time of each AMR and generate collaborative instructions. The relevant AMRs execute these instructions to resolve the deadlock. If the AMR does not receive an arbitration instruction within the preset timeout period after sending a deadlock report to the edge node, it automatically degrades to an autonomous obstacle avoidance mode that determines the yielding order based on the priority of the local cache. After the AMR reaches the target workstation, it performs the handling and docking action, and reports task completion after confirming that the handling is stable through limit switches and inertial measurement units.

[0091] Example 4:

[0092] In this embodiment, as the AMR travels along the global reference path, the onboard LiDAR and depth camera continuously collect environmental data. The processor fuses the point cloud and images to detect objects with a speed exceeding a preset threshold. The visual classification model identifies the detected objects and outputs classification results for pedestrians, other AMRs, or temporary materials.

[0093] When an obstacle is identified as a pedestrian, the processor invokes a motion prediction model that combines a constant velocity model with social force correction. The constant velocity model calculates the future position sequence based on the pedestrian's current linear velocity and heading angle. The social force correction module applies a virtual repulsive force during trajectory generation. This repulsive force is directed from the pedestrian towards other surrounding targets, and its magnitude is inversely proportional to the distance, causing the predicted trajectory to simulate the pedestrian's tendency to maintain distance from other targets.

[0094] When an obstacle is identified as another AMR, this AMR receives the current target path information broadcast by that AMR via the onboard V2X communication module and extracts its linear velocity and angular velocity values. The processor calls the constant curvature and constant acceleration motion model, using the received linear velocity and angular velocity as the initial motion state, and extrapolates the trajectory along the target path to generate the predicted trajectory of the AMR in the future time period.

[0095] When an obstacle is identified as temporary material, the processor marks the material as a stationary obstacle with a fixed position in the dynamic cost map. The grid it occupies remains constant during the prediction period and no motion trajectory is generated.

[0096] Example 5:

[0097] In this embodiment, the multi-AMR cooperative transport and obstacle avoidance method in dynamic scenarios receives a transport task, assigns tasks to each AMR, and sets an initial priority. Each AMR constructs a dynamic cost map based on a static map overlaid with a heatmap generated from historical traffic data and maintained using exponential decay. The cost function of the path search algorithm incorporates a penalty term from the heatmap to plan a global reference path. The AMR travels along the global reference path, perceiving the environment in real time through onboard LiDAR and depth cameras. It detects objects with a speed exceeding a preset threshold, identifies obstacle types using a visual classification model (including at least pedestrians, other AMRs, and temporary materials), and generates predicted trajectories based on the corresponding motion prediction model according to different obstacle types.

[0098] Spatiotemporal conflict detection is performed based on the planned path of this AMR and the predicted trajectory of obstacles. When a conflict involves other AMRs, the conflict area is divided into multiple resource blocks, and the planning order and time window of each related AMR entering each resource block are obtained. A directed graph is constructed with AMRs as nodes and resource occupation waiting relationships as edges. If a resource block planned by one AMR is reserved by another AMR and the occupation time periods of the two overlap, an edge is added from the first AMR to the second AMR. A depth-first search algorithm is run on the constructed directed graph to detect loops. If a loop exists, it is identified as a deadlock risk, and the AMRs included in the loop are treated as participants in the cooperative avoidance strategy, triggering the edge-end cooperative avoidance strategy. The edge node arbitrates based on the priority and waiting time of each AMR and generates a cooperative instruction. The relevant AMRs execute the cooperative instruction to resolve the deadlock. If an AMR does not receive an arbitration instruction within a preset timeout after sending a deadlock report to the edge node, it automatically degrades to an autonomous obstacle avoidance mode that determines the yielding order based on the priority of its local cache. If there is no loop but an edge exists, it is considered a general conflict that is not deadlocked. The AMR will autonomously perform local path replanning and decelerate to detour. The local path replanning is guided by the global reference path. After the AMR reaches the target workstation, it will perform the handling and docking operation. After confirming that the handling is stable through limit switches and inertial measurement units, it will report the task as completed.

[0099] Example 6:

[0100] In this embodiment, the AMR onboard controller obtains a global reference path based on a dynamic cost map from the global planner. This global reference path actively avoids historical congestion areas during planning by using a heat matrix penalty term in the cost function. As the AMR travels along the global reference path, the onboard LiDAR and depth camera continuously collect environmental data. When the spatiotemporal conflict detection module of this AMR determines that the current conflict is a general conflict that is not deadlocked, a local path replanning process is triggered.

[0101] Local path replanning uses the global reference path as a guiding benchmark, and the specific execution process is as follows: The AMR onboard controller calls the dynamic window method module to perform discrete sampling in a two-dimensional velocity space composed of linear velocity and angular velocity, generating multiple sets of velocity command pairs. For each set of velocity command pairs, the dynamic window method module simulates and generates a candidate trajectory forward within a preset time window. The evaluation function module receives all candidate trajectories and calculates the comprehensive evaluation value of each trajectory. The evaluation function includes a progress component toward the global path, which forces the global reference path as the calculation benchmark. Specifically, it calculates the directional deviation and lateral distance deviation between the pose of the candidate trajectory end and the corresponding nearest point on the global reference path, and maps the deviation to the progress component score. Since the global reference path itself has bypassed the historical congestion areas marked by the heat matrix, the forced guidance of the progress component makes the local detour trajectory tend to extend along the direction of the global reference path while avoiding the current dynamic obstacles, thereby preventing the AMR from entering historically high-frequency congested channels or intersections due to temporary detours.

[0102] The evaluation function module calculates the total score for each candidate trajectory by comprehensively considering factors such as progress, obstacle distance, and speed, and selects the candidate trajectory with the highest total score as the optimal trajectory. The AMR onboard controller sends the speed command corresponding to the optimal trajectory to the chassis drive module for execution. Simultaneously, the AMR onboard controller reduces the driving speed limit from the current setting to a preset safety value, which is stored in the AMR's local parameter configuration table. The chassis drive module responds to the speed reduction command, controlling the drive motor and steering motor to ensure the AMR travels along the optimal trajectory at a speed not exceeding the preset safety value. The AMR continuously performs spatiotemporal conflict detection. When it is detected that the AMR has completely left the conflict area, the onboard controller restores the driving speed limit to the original setting, and the AMR continues to travel towards the target workstation along the global reference path.

[0103] Example 7:

[0104] In this embodiment, the edge node continuously receives status information reported by each AMR. This status information includes the AMR identifier, current task priority value, current location coordinates, current velocity vector, and the planned local path. When the edge node receives deadlock reports from at least two AMRs, it extracts the conflict area range and the identifiers of each AMR involved in the deadlock from the reports. It then queries the task priority value of each AMR from the status information and identifies the AMR with the lowest priority value as the priority traveler. If two or more AMRs have the same priority value, the edge node extracts the accumulated waiting time of each AMR within the conflict area from the status information reported by each AMR. This accumulated waiting time is calculated from the moment the AMR enters the edge of the conflict area and continues to be counted. The edge node compares the accumulated waiting times of AMRs with the same priority and identifies the AMR with the longest accumulated waiting time as the priority traveler. The edge node generates a cooperative instruction for the priority traveler, including a command to continue along the currently planned path, and issues this cooperative instruction to the priority traveler. For an AMR identified as needing to yield, the edge node obtains the current coordinates of the AMR and checks if there are other AMRs behind it. The determination of whether other AMRs are behind it is based on whether any other AMR markers are detected within a preset distance range behind the AMR's current lane. If no other AMRs are behind it, the edge node generates a coordinated instruction containing a pause-yield command. This pause-yield command instructs the AMR to stop at its current location and remain stationary until it receives a resume-traffic command from the edge node. If other AMRs are behind it, the edge node extracts the coordinates of the nearest yield point behind the AMR's current location from the static map. The yield point is a pre-marked temporary parking area on the static map. The edge node generates a coordinated instruction containing a reverse-to-yield command. This reverse-to-yield command includes the target yield point coordinates and the reverse path. The edge node then sends this coordinated instruction to the AMR that needs to yield. After receiving the coordination instruction, the AMR that needs to give way will either stop and wait or reverse along the back path to the target avoidance point, depending on the instruction type, and report the arrival confirmation information to the edge node after reaching the avoidance point.

[0105] Example 8:

[0106] In this embodiment, the edge node receives deadlock reports from each AMR and extracts the task priority value and accumulated waiting time in the conflict area for each AMR involved in the deadlock. The edge node compares the task priority values ​​of each AMR and marks the AMR with the lowest priority value as the priority AMR. When two or more AMRs have the same task priority value, the edge node further compares the accumulated waiting time of these AMRs with the same priority and marks the AMR with the longest accumulated waiting time as the priority AMR. The edge node generates a yielding coordination instruction for AMRs that have not obtained priority passage. The generation process is as follows: the edge node obtains the current position coordinates of the AMR and queries whether there are other AMRs within a preset distance range behind it. If there are no other AMRs behind it, the edge node generates a pause yielding instruction, which includes the AMR's identifier, the current path node number, and the waiting time in place. If other AMRs are located behind, the edge node extracts the coordinates of the nearest avoidance point behind the current position of the AMR from the global reference path and generates a retreat to avoidance point instruction. This instruction includes the AMR's identifier, the target avoidance point coordinates, and the retreat speed. After generating the yield coordination instruction, the edge node checks whether the task priority of the yielding AMR is lower than a preset threshold and determines whether the spatial size of the conflict area is less than the sum of the minimum accommodating areas required for all AMRs that need to be paused. When the task priority of the yielding AMR is lower than the preset threshold and the conflict area space is insufficient to accommodate all AMRs that need to be paused, the edge node cancels the generated pause yield instruction or retreat to avoidance point instruction and instead issues a global path replanning instruction to the low-priority AMR. The global path replanning instruction includes the AMR's identifier, the current node, the target workstation node, and the avoidance area boundary coordinate set, which is generated by expanding the conflict area by a preset safety distance. After receiving the global path replanning instruction, the onboard controller of the low-priority AMR marks the avoidance area as an impassable grid on the dynamic cost map, re-executes the path search with the current position as the starting point and the target workstation as the ending point, generates a new global reference path that avoids the current conflict area, and continues to travel along the new global reference path.

[0107] Example 9:

[0108] In this embodiment, during the collaborative transport task execution, if multiple AMRs form a deadlock in the conflict area and each AMR has sent a deadlock report to the edge node, and no arbitration instruction is received from the edge node within a preset timeout period, the autonomous obstacle avoidance mode degradation process is triggered. The control units of each AMR involved in the deadlock read pre-cached task priority information from their local storage modules. This priority information is written to the memory of each AMR by the edge node during the task allocation phase and sent along with the task data. The AMR control unit transmits the read local priority value to the decision module via the internal bus. The decision module compares the priority values ​​of each AMR involved in the conflict. The AMR with the lower priority value is determined to have higher right-of-way, and its motion control unit maintains its current speed and path, continuing to pass through the conflict area. The AMR with the higher priority value is determined to need to yield; its motion control unit calculates a deceleration curve based on the distance to the conflict area and its current speed, driving the walking mechanism to perform a deceleration action, or pausing and waiting at the entrance of the conflict area until the AMR with higher right-of-way completely leaves the conflict area. If the decision module finds that the priority values ​​in the local cache are the same, it further reads the device identifier of each AMR. The device identifier is a unique code burned into the control unit firmware at the factory. The decision module sorts the device identifiers in ascending or descending order according to the character sequence. The AMR with higher priority is granted the right to pass through the conflict area first, while the AMR with lower priority is ordered to slow down and give way or pause and wait. Each AMR continuously monitors the movement status of other AMRs in the conflict area through the vehicle-mounted LiDAR and depth camera. When it is detected that the AMR with higher priority has left the conflict area, the control unit of the AMR that gave way resumes driving and continues to perform the transport task along the global reference path.

[0109] Example 10:

[0110] In this embodiment, the AMR vehicle-mounted embedded computing unit is connected to the depth camera via a PCIe interface and to the LiDAR via an Ethernet interface. The LiDAR emits a laser beam to scan the environment and receives the reflected point cloud. The embedded computing unit runs a point cloud clustering algorithm to segment the point cloud into independent point clusters, with each cluster corresponding to a potential obstacle. The depth camera simultaneously acquires RGB images, and the image data is transmitted to the embedded computing unit via a USB 3.0 bus.

[0111] The feature fusion module within the embedded computing unit projects the clustered point clouds from the LiDAR onto the pixel coordinate system of the RGB image from the depth camera. It extracts geometric features from each point cluster and texture features from the corresponding image region. The geometric and texture features are then concatenated along the channel dimension to form a fused feature map. This fused feature map is input to a lightweight object detection and instance segmentation model deployed on the embedded computing unit. This model is based on the MobileNetV3 backbone network and includes a feature pyramid network and a detection head.

[0112] The model performs multi-scale feature extraction on the fused feature map. The detection head outputs the bounding box coordinates and class confidence of obstacles, while the instance segmentation branch outputs pixel-level masks of obstacles. The model compares the class confidence with preset class labels; when the confidence exceeds 0.7, the obstacle type is determined to be one of pedestrians, other AMRs, or temporary materials. The model combines the bounding box coordinates and mask regions to generate the obstacle's contour information, represented by a sequence of polygon vertices. The embedded computing unit writes the obstacle type and contour information to the AMR main control unit via shared memory for use by the motion prediction model.

[0113] Example 11:

[0114] In this embodiment, after the AMR reaches the target workstation, it performs a transport and docking operation. The lifting mechanism moves upward from inside the AMR body, and the trigger plate at the bottom of the lifting mechanism moves synchronously with the lifting mechanism. The limit switch is fixedly installed at a preset height position inside the AMR body, with the detection end of the limit switch facing the movement path of the trigger plate. When the lifting mechanism reaches the target position, the trigger plate enters the detection range of the limit switch, the internal contacts of the limit switch close, and a first confirmation signal is output to the AMR's on-board controller.

[0115] The inertial measurement unit (IMU) is rigidly fixed to the upper surface of the AMR's cargo platform, with its measurement axis aligned with the cargo platform's coordinate system. The IMU continuously monitors the cargo platform's current roll and pitch angles and transmits these values ​​to the onboard controller in real time. The onboard controller compares the received roll and pitch angle values ​​with preset threshold values. When both the roll and pitch angle values ​​do not exceed the preset threshold values, the onboard controller determines that the transport attitude is stable.

[0116] Upon receiving the first confirmation signal and determining that the transport posture is stable, the vehicle controller generates a locking command. This locking command is sent to the drive motor controller of the lifting mechanism, which stops outputting drive current to the lifting mechanism. The lifting mechanism maintains its current position, and the loading state is locked. The vehicle controller then sends a task completion message to the warehouse management system or manufacturing execution system via its wireless communication module. This message includes the AMR number, task number, and timestamp.

[0117] Example 12:

[0118] The scheduling system in this embodiment receives handling task messages from the warehouse management system via a wireless communication network. These messages include the task number, starting workstation coordinates, ending workstation coordinates, and material attributes. The scheduling system queries the current status of each AMR (Autonomous Mobile Transporter), including its location, battery level, cargo status, and current task queue length. Based on a load balancing strategy, tasks are assigned to each AMR, and an initial priority value is set according to the task urgency and material attributes. The priority value is an integer from 0 to 255, with smaller values ​​indicating higher priority. The initial priority is written to the onboard non-volatile memory of each AMR.

[0119] Each AMR's onboard computing unit reads a static raster map from the storage module. The static raster map has a resolution of 0.05 meters per grid, and each grid stores its occupancy status. The onboard computing unit reads a heat map from the historical passage data storage module. Each grid in the heat map records the cumulative number of times the AMR has passed through in the past 600 seconds. The cumulative count is based on the grid where the AMR's center point is located, and the count of the corresponding grid is incremented by 1 each time the AMR passes through. Every 30 seconds, the onboard computing unit triggers an exponential decay process. The decay process traverses all grids in the heat map, multiplies the old heat map value of each grid by a decay coefficient of 0.5, adds the new passage count added during the current 30-second statistical period, and obtains a new heat map value, which is then written back to the heat map matrix. The onboard computing unit overlays the heat map matrix onto the static raster map to construct a dynamic cost map. The overlay method involves adding a heat cost value field to each grid of the static map.

[0120] The path search algorithm uses the A* algorithm. When expanding the neighboring nodes of the current node, the cost function f(n) = g(n) + h(n) + α × HeatMap[n] × dist_factor is calculated, where g(n) is the actual cumulative Euclidean distance from the starting point to the current node n along the searched path, h(n) is the heuristic value of the Euclidean distance from the current node n to the target node, α is 2.0, HeatMap[n] is the heat value of the raster corresponding to node n, dist_factor = 1.0 / (1.0 + exp(-5.0 × (d / D - 0.5))), d is the Euclidean distance from node n to the target node, and D is the Euclidean distance from the starting point to the target node. After the path search is completed, a global reference path is generated and stored in the AMR local memory in the form of a raster coordinate sequence.

[0121] During the AMR's journey along the global reference path, the onboard LiDAR emits a 905nm wavelength laser beam at a 10Hz frequency and receives reflected point clouds, while the depth camera acquires 640×480 resolution depth and RGB images at a 30Hz frequency. The onboard computing unit performs hardware timestamp alignment on the LiDAR point clouds and depth camera images, matching the depth camera image timestamp with the nearest neighbor LiDAR point cloud timestamp within a 5ms window. The onboard computing unit performs Euclidean clustering on the point clouds, with a clustering distance threshold of 0.1 meters, generating independent point clusters. For each point cluster, the onboard computing unit calculates a 3D bounding box, extracting the bounding box's length, width, and height dimensions, as well as the cluster center coordinates, as geometric features. The onboard computing unit projects the 3D bounding boxes of the point clusters onto the RGB image plane, using a joint calibration extrinsic parameter matrix from the LiDAR and depth camera. The projected area is cropped from the RGB image, and HOG texture feature vectors are extracted. The onboard computing unit concatenates the geometric feature vectors with the HOG texture feature vectors along the feature dimension, forming a 128-dimensional fused feature vector.

[0122] A lightweight object detection and instance segmentation model, integrating feature vector inputs and deployed on an in-vehicle embedded computing unit, is proposed. This model employs a MobileNetV3-Small backbone network, followed by a feature pyramid network and a dual-branch detection head. The classification branch of the detection head outputs obstacle category probability vectors, including pedestrians, other AMRs, and temporary materials. The instance segmentation branch outputs binary masks of obstacles. When the maximum category probability exceeds 0.7, the model outputs the corresponding category label and mask contour, with the mask contour represented by a sequence of polygon vertices.

[0123] When the obstacle type is a pedestrian, the onboard computing unit calls a motion prediction model combining a constant velocity model and a social force correction. The constant velocity model takes the pedestrian's current linear velocity and heading angle as input, extrapolates the position sequence within 5 seconds in 0.1-second time steps. The social force correction term applies a repulsive acceleration to the pedestrian's position at each time step. The direction of the repulsive acceleration is from the pedestrian towards other surrounding targets, and its magnitude is equal to 0.5 divided by the square of the distance from the pedestrian to the target. When the obstacle type is another AMR, the onboard computing unit receives the current target path information, linear velocity, and angular velocity broadcast by the AMR at a frequency of 10 Hz via the V2X communication module. It calls a constant curvature and constant acceleration motion model, using the linear velocity and angular velocity at the time of reception as the initial state, and extrapolates the predicted trajectory within 5 seconds along the target path in 0.1-second time steps. When the obstacle type is temporary material, the onboard computing unit marks the material as a static obstacle in the dynamic cost map, and the occupied grid remains unchanged during the prediction period.

[0124] The onboard computing unit discretizes the global reference path of this AMR into a spatiotemporal trajectory point sequence with a time step of 0.1 seconds. Each trajectory point contains a timestamp, grid coordinates, and the 2D bounding box of this AMR. The onboard computing unit performs spatiotemporal conflict detection between the spatiotemporal trajectory point sequence of this AMR and the predicted trajectories of each obstacle. The conflict determination condition is that the 2D bounding box of this AMR and the 2D bounding box of the obstacle intersect at the same timestamp. When the conflict involves other AMRs, the onboard computing unit divides the conflict area into multiple resource blocks using a 0.5m × 0.5m grid, with the resource blocks identified by row and column indices. The onboard computing unit obtains the planned time window for each relevant AMR to enter each resource block, with the time window represented by the entry time and departure time.

[0125] The onboard computing unit constructs a directed graph, stored using an adjacency list data structure. Nodes are identified by AMRs (Adjacent Registries), and edges represent resource occupancy and waiting relationships. The onboard computing unit traverses all resource blocks. For each resource block, it compares the occupancy time windows of each AMR. If the occupancy time windows of AMR-A and AMR-B overlap, and AMR-A's entry time is later than AMR-B's, a directed edge from node A to node B is added. The onboard computing unit runs a depth-first search algorithm on the directed graph to detect loops. The algorithm maintains three state arrays: unvisited, searching, and completed. The search recursively starts from each unvisited node. If a node in the searching state is encountered, the path from that node to the current node is recorded as a loop, indicating a deadlock risk. The list of AMRs contained in the loop is used as participants in the collaborative avoidance mechanism. If the search is complete and no loop is found, but a directed edge exists, it is considered a general conflict that is not deadlock-prone.

[0126] When a non-deadlock general conflict is identified, the onboard computing unit triggers local path replanning. The onboard computing unit invokes the dynamic window method module to sample within the velocity space with a linear velocity resolution of 0.05 m / s and an angular velocity resolution of 0.1 radians / s, where the linear velocity ranges from 0 to 1.0 m / s and the angular velocity ranges from -1.0 to 1.0 radians / s. For each sampled velocity, the dynamic window method module simulates and generates candidate trajectories within a 3-second time window with a step size of 0.1 seconds. The evaluation function module calculates the total score for each candidate trajectory: Total Score = 0.4 × Progress Component + 0.3 × Obstacle Distance Component + 0.3 × Velocity Component. The progress component is calculated by finding the path point on the global reference path with the closest Euclidean distance to the candidate trajectory's end pose, and calculating the angular deviation between the candidate trajectory's end heading angle and the tangent direction of that path point. The progress component score = 1.0 / (1.0 + Angular Deviation + Lateral Distance Deviation). The onboard computing unit selects the candidate trajectory with the highest total score as the optimal trajectory, sends the speed command corresponding to the optimal trajectory to the chassis drive module, and reduces the driving speed limit from 1.0 m / s to 0.3 m / s until the onboard computing unit detects that the intersection of the two-dimensional bounding box of this AMR and the conflict area is empty, and then restores the original speed limit.

[0127] When a deadlock risk is detected, the AMR's communication module sends a deadlock report to the edge node via the 5G wireless network. The deadlock report includes the AMR's identifier, a list of resource block indexes in the conflict area, a list of all AMR identifiers in the loop, and the AMR's accumulated waiting time in the conflict area. Upon receiving the deadlock report, the edge node's arbitration module extracts the task priority values ​​of each AMR involved in the deadlock and determines the AMR with the lowest priority value as the priority AMR. If AMRs with the same priority value exist, the edge node compares their accumulated waiting time, calculated from the moment the AMR first enters the conflict area edge. The AMR with the longest accumulated waiting time is determined as the priority AMR. The edge node generates a continue driving instruction for the priority AMR and a yield instruction for AMRs that need to yield. When a yield command is generated, the edge node checks if there are other AMRs within 5 meters behind the current AMR's location. If not, it generates a stop command with a maximum pause duration. If other AMRs exist, it extracts the coordinates of the nearest avoidance point behind the current AMR's location from the static map. The avoidance point is a pre-marked 2m x 2m clear area. A backtrack command is generated, containing the target avoidance point coordinates and a backtracking speed of 0.2 meters per second. If the AMR requiring yield has a task priority value below a preset threshold of 50 and the conflict area is smaller than the sum of the minimum areas required for all AMRs to pause, the edge node cancels the yield command and issues a global path replanning command to the AMR. This command contains the boundary coordinate set of the avoidance area, which is generated by expanding the conflict area by 1 meter. After receiving the global path replanning command, the AMR's onboard computing unit marks the avoidance area grid as impassable on the dynamic cost map. Starting from the current location, it re-executes the A* path search, generates a new global reference path, and travels along the new path.

[0128] If the AMR does not receive an arbitration command from the edge node within 500 milliseconds after sending a deadlock report, the onboard computing unit triggers the autonomous obstacle avoidance mode degradation process. The onboard computing unit reads the task priority value of this AMR from its local non-volatile memory and broadcasts the priority value and device identifier of this AMR at a frequency of 10 Hz via the V2X communication module, while simultaneously receiving the priority values ​​and device identifiers broadcast by other AMRs in the conflict area. The onboard computing unit compares the priority values ​​of all participating AMRs. If the priority value of this AMR is the lowest, it maintains its current driving state and passes through the conflict area; if the priority value of this AMR is higher, it performs deceleration and yielding, with the deceleration curve being a linear reduction in speed to 0 within 1 second and maintaining a stationary position; if the priority values ​​are the same, they are sorted in ascending lexicographical order by device identifier, with the AMR ranked earlier having priority to pass and the AMR ranked later performing yielding.

[0129] After the AMR reaches the target workstation, the on-board controller outputs a lifting command to the lifting mechanism drive motor, and the lifting mechanism moves upward at a speed of 0.05 meters per second. The limit switch is a mechanical contact type, fixedly installed inside the AMR body at a height of 0.3 meters above the ground, and the trigger plate at the bottom of the lifting mechanism is a metal baffle. When the lifting mechanism rises to the target position, the trigger plate enters the detection range of the limit switch, the internal contacts of the limit switch close, and the first confirmation signal is output to the on-board controller.

[0130] The inertial measurement unit (IMU) is rigidly fixed to the upper surface of the AMR's cargo platform, with its measurement axis aligned with the cargo platform's coordinate system. It continuously monitors the roll and pitch angle values ​​of the cargo platform and transmits them to the vehicle controller in real time. The vehicle controller compares the roll angle value with a preset roll angle threshold and the pitch angle value with a preset pitch angle threshold. When both the roll angle and pitch angle do not exceed the preset roll angle threshold, the handling posture is determined to be stable.

[0131] Upon receiving the first confirmation signal and determining that the transport posture is stable, the vehicle-mounted controller generates a locking command and sends it to the lifting mechanism drive motor controller. The drive motor controller stops outputting drive current, the lifting mechanism maintains its current position, and the loading state is locked. Subsequently, the vehicle-mounted controller sends a task completion message to the warehouse management system or manufacturing execution system via the wireless communication module. The message includes the AMR number, task number, and timestamp, indicating that the task is completed and reported.

[0132] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0133] The above description is only used to illustrate the technical solution of the present invention and is not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A multi-AMR cooperative transport obstacle avoidance method in dynamic scenarios, characterized in that, Includes the following steps: Receive handling tasks, assign tasks to each AMR and set initial priorities; For each AMR, a dynamic cost map is constructed based on a static map overlaid with a heat matrix generated from historical traffic data and maintained using exponential decay. The cost function of the path search algorithm introduces a penalty term into the heat matrix to plan a global reference path that can proactively avoid historical congestion areas. Each AMR travels along the global reference path and simultaneously perceives the environment in real time through onboard LiDAR and depth camera. It detects objects with a speed higher than a preset threshold and identifies obstacle types through a visual classification model. The types include at least pedestrians, other AMRs, and temporary materials. Based on different obstacle types, it calls the corresponding motion prediction model to generate predicted trajectories. Spatiotemporal conflict detection is performed based on the planned path and predicted trajectory of obstacles of this AMR. When the conflict involves other AMRs, the conflict area is divided into multiple resource blocks. The planning order and time window of each related AMR entering each resource block are obtained. A directed graph with AMRs as nodes and resource occupation waiting relationship as edges is constructed. The presence of deadlock risk is determined by detecting loops on the directed graph. If only general conflicts that are not deadlocked exist, the AMR will autonomously perform local path replanning and decelerate to detour. The local path replanning is guided by the global reference path. If a deadlock risk exists, an edge-to-edge collaborative obstacle avoidance strategy is triggered. The edge node arbitrates based on the priority and waiting time of each AMR and generates a collaborative instruction. The relevant AMR executes the collaborative instruction to resolve the deadlock. If an AMR does not receive an arbitration instruction within a preset timeout after sending a deadlock report to the edge node, it automatically degrades to an autonomous obstacle avoidance mode that determines the yielding order based on the priority of its local cache. Once the AMR reaches the target workstation, it performs the handling and docking action, and after confirming the stability of the handling through limit switches and inertial measurement units, it reports the task as completed.

2. The multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The maintenance methods for the thermal matrix include: Each grid records the cumulative number of times the AMR passes through within a preset time period, and performs exponential decay processing according to a preset period. The new thermal value is the old thermal value multiplied by the decay coefficient of 0.5 and then added to the count added in the current statistical period. The cost function is f(n)=g(n)+h(n)+α×HeatMap[n]×dist_factor, where α is the weight of thermal influence and dist_factor is the attenuation factor that varies with the distance from the node to the target point, so as to avoid excessive detours on road segments near the end of the path.

3. The multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The step of generating predicted trajectories by calling the corresponding motion prediction model according to different obstacle types includes: For pedestrians, a constant velocity model combined with social force correction is used for trajectory prediction. The social force correction is used to simulate the pedestrian's tendency to keep a distance from other targets. For other AMRs, their current target path information is obtained through V2X communication, and their linear velocity and angular velocity are combined to perform trajectory extrapolation using a motion model with constant curvature and constant acceleration. Temporary materials are treated as stationary obstacles with fixed positions.

4. The multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The construction of a directed graph with AMRs as nodes and resource occupancy / wait relationships as edges, and the determination of deadlock risk by detecting cycles in the directed graph, includes: If a resource block planned to be occupied by one AMR is reserved by another AMR and the occupation time periods of the two overlap, then add an edge from the first AMR to the other AMR. Run a depth-first search algorithm on the constructed directed graph to detect loops. If a loop exists, it is considered a deadlock risk, and the AMRs contained in the loop are identified as participants in the cooperative avoidance. If there are no cycles but there are edges, it is considered a general conflict that is not deadlocked.

5. The multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The process of autonomously performing local path replanning and decelerating around obstacles by this AMR, wherein the local path replanning is guided by the global reference path, includes: Candidate trajectories are generated by sampling in the velocity space using the dynamic window method. The progress component toward the global path in the evaluation function is forced to be guided by the global reference path planned based on the dynamic cost map, so that the local detour trajectory tends to avoid historical congestion areas while avoiding obstacles. Select the optimal trajectory and reduce the AMR's speed limit from the current value to a preset safe value until it safely passes through the conflict area and then resumes the original speed.

6. The multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The process of arbitrating and generating collaborative instructions by edge nodes based on the priority and waiting time of each AMR includes: Compare the task priorities of each AMR involved in the deadlock, and the task with the lowest priority value will have priority to pass. If there are AMRs with the same priority, then the accumulated waiting time of each AMR in the conflict area is further compared, and the AMR with the longer waiting time is given priority to pass. For AMRs that need to yield, an instruction is generated to pause yielding or back up to the avoidance point based on their current position and whether there are other AMRs behind them.

7. A multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 6, characterized in that, The arbitration and generation of collaborative instructions also includes: If there is an AMR with a task priority lower than the preset threshold and the conflict area space is insufficient to accommodate all AMRs that need to be paused, then a global path replanning instruction is directly issued to the low-priority AMR, so that it actively avoids the current conflict area.

8. The multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The automatic degradation is an autonomous obstacle avoidance mode that determines the yielding order based on the priority of the local cache, including: If each AMR involved in deadlock does not receive an arbitration instruction from the edge node within a preset timeout period, it will read the priority information from its local cache. Those with lower priority values ​​pass through the conflict zone first, while those with higher priority values ​​either slow down to give way or pause and wait. If the local caches have the same priority, the passage order is determined by sorting the device identifiers in the AMR.

9. A multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The visual classification model is a lightweight target detection and instance segmentation model deployed on an AMR vehicle-mounted embedded computing unit. The model receives the feature fusion results of point clusters after LiDAR point cloud clustering and RGB images of depth cameras as input, and outputs the type and contour information of obstacles.

10. A multi-AMR cooperative transport and obstacle avoidance method in a dynamic scene according to claim 1, characterized in that, The step of reporting task completion after confirming stable handling via limit switches and inertial measurement units includes: Confirm that the lifting mechanism has reached the target position by using the limit switch; The roll or pitch angle deviation of the current attitude is detected by the inertial measurement unit and does not exceed a preset threshold. Once the above conditions are met, the loading status is locked, and then the task is reported as completed to the warehouse management system or manufacturing execution system.

Images