Path planning method and system for multiple mobile robots

Abstract

The invention discloses a path planning method and system for multiple mobile robots, and the method comprises the steps: 1, constructing a real-time map through the fusion of multiple sensors, and marking the track of an obstacle; the edge node lightweight network identifies the type of an obstacle, the LSTM predicts the future position of the obstacle, and a high-risk target is preferentially processed in combination with a risk model; 2, dynamically clustering, and managing intra-cluster cooperation by edge nodes; an A * algorithm is improved to generate candidate paths, a Pareto optimal path set is screened, and cluster scale balance calculation and communication efficiency are adaptively adjusted; 3, dynamically adjusting weights of path length, energy consumption and safety based on the environmental risk value, introducing a grey wolf optimizer to generate a Pareto leading edge path, and balancing multiple targets in real time; 4, the MCTS predicts the path conflict probability and triggers a distributed auction mechanism; robot bidding and priority negotiation of high-priority tasks are carried out, and conflicts are solved in advance; and step 5, multiple sensors feed back and correct path deviation, and historical data are stored in an experience pool training element learning model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of path planning, in particular to a path planning method and system for multiple mobile robots. BACKGROUND

[0002] Multiple mobile robots have been widely used in warehouse logistics, medical distribution, intelligent manufacturing and other fields, such as robot clusters carrying goods in e-commerce warehouses, distribution robots delivering medical supplies in hospital corridors, and AGVs transporting parts in intelligent manufacturing workshops. In all these scenarios, robots need to efficiently and safely complete complex tasks.

[0003] However, current mainstream path planning methods mostly rely on centralized control or fixed rule distributed algorithms, which have significant adaptability problems in dynamic environments. For example, when a forklift suddenly enters the robot work area in a warehouse scenario, the traditional method may cause collision due to time-consuming global path re-planning. In a hospital scenario, a robot that urgently transports emergency medicines may be blocked due to the lack of dynamic adjustment of priority by ordinary task robots, delaying treatment. Therefore, a path planning method and system for multiple mobile robots are proposed. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art, and a path planning method and system for multiple mobile robots are proposed.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme: The path planning method for multiple mobile robots comprises: Step 1: Real-time construction of environment map through multi-sensor fusion (laser radar, camera, IMU), labeling the motion trajectory of dynamic obstacles (such as moving targets, sudden obstacles); using a lightweight neural network deployed on the edge node to quickly identify the type of obstacle (static / dynamic), and predicting its future position through LSTM; combining a dynamic obstacle risk level evaluation model, calculating the collision risk value based on speed, direction and size, and prioritizing high-risk targets to provide real-time and accurate dynamic environment information for subsequent path planning; Step 2: Dynamically dividing clusters according to robot position, task priority and communication range, with each cluster managed by an edge node (such as a roadside unit); robots within a cluster share local maps and path intentions, and the edge node generates candidate paths through an improved A* algorithm, and selects a Pareto optimal path set based on non-dominated sorting and congestion comparison; adaptively adjusting the cluster size based on environmental complexity (such as obstacle density) to balance computational load and communication efficiency, and fusing local real-time decision-making and global coordination; Step three: dynamically adjust the weights of path length, energy consumption, and safety based on environmental risk values to avoid suboptimal solutions caused by fixed weights; introduce an improved multi-objective grey wolf optimizer, combine non-dominated sorting and congestion comparison to generate a Pareto frontier path; through the weight self-adaptive mechanism of environmental perception, real-time balance multi-objective, such as emergency task priority to shorten the path, low energy consumption task priority to save energy path, improve the comprehensive optimization ability in dynamic scene; Step four: based on the current path of the robot and the trajectory of the dynamic obstacle, predict the future conflict probability through Monte Carlo tree search; use a distributed auction mechanism, and the robot generates a bid according to the path cost (remaining distance, energy consumption) and task priority; high-priority tasks (such as medical supply transportation) have higher bidding weights; combine task urgency and robot remaining energy to dynamically adjust the bidding weight, predict conflicts 1-2 steps in advance and negotiate solutions, reduce the number of real-time negotiations, and reduce communication delay and conflict rate; Step five: the robot corrects the path deviation in real time through the odometer and visual odometry, and triggers local re-planning if the deviation exceeds the threshold; store historical dynamic scene data (obstacle movement pattern, conflict resolution case) into the experience replay pool to train the meta-learning model; when encountering a new scene, the model quickly generates an adaptive strategy through a small amount of gradient updates, performs "zero-shot" or "small-shot" scene adaptation, breaks through the dependence of traditional deep learning on training data, and improves the generalization ability and real-time performance in dynamic environments.

[0006] The above technical solutions further include: Further, the light neural network deployed by the edge node quickly identifies the type of obstacle and predicts its future position through LSTM; combine the dynamic obstacle risk level evaluation model to prioritize high-risk targets, including the following steps: Lightweight neural network obstacle type identification: The edge node (such as a robot local or roadside unit) receives multi-sensor fusion data (laser radar point cloud, camera image, IMU attitude information), forms an original data stream containing obstacle geometric features and motion state, and inputs a lightweight neural network; the lightweight neural network refers to compressing the parameters of MobileNetV3 model to 1 / 4 of the original size by 8-bit integerization, using structured pruning to remove redundant convolution kernels and unstructured pruning to remove low-weight connections, and introducing depth separable convolution to decompose standard convolution into depth convolution and point convolution, finally obtaining an improved MobileNetV3 lightweight model, extracting the visual features and motion features of the obstacle through the convolution layer, and finally determining the obstacle type through the classification layer, achieving a reasoning speed of 10-20 ms on embedded devices; LSTM network trajectory prediction: Using the position coordinates of dynamic obstacles over the past 3-5 seconds, such as 2D / 3D points detected by LiDAR, as the input sequence, a time window data is constructed. Through the hidden state memory mechanism of the LSTM unit, the motion pattern of the obstacle is learned, such as uniform linear motion, acceleration, and turning, and the predicted trajectory for the next 1-2 seconds is output, represented in the form of a probability distribution, such as a Gaussian process or particle filter result. The confidence score of the predicted trajectory is performed. If the confidence score is lower than the threshold, sensor resampling or fusion of other sensor data, such as radar speed information, is triggered to improve the reliability of the prediction. Dynamic risk level assessment: Risk indicator calculation: Speed ​​risk: Radial velocity of the obstacle relative to the robot With threshold ratio =+ Obstacles approaching at high speed pose a higher risk; Directional risk: The cosine of the angle θ between the obstacle's direction of motion and the robot's path. Obstacles oriented towards the robot's path pose a higher risk; Size risk: The volume V of the space occupied by the obstacle and the safety threshold ratio Large obstacles pose a higher risk; By weighted summation Alternatively, fuzzy logic can be used to fuse the above indicators to generate a risk value of 0-100. Based on the risk value ranking, obstacles with a risk value greater than 80 (such as large vehicles approaching at high speed) will immediately trigger local path replanning and adjust the robot's movement speed, such as slowing down or detouring. For obstacles with a risk value less than 50, such as slowly moving pedestrians, monitoring will be maintained, and path planning will not be intervened for the time being. The identified obstacle types, predicted trajectories, risk values, and priority markers are updated in real time to the global environment map and broadcast to other robots in the cluster (via edge nodes). Information on high-risk targets is directly input into the dynamic clustering module in step two, affecting the path generation and coordination strategies within the cluster, such as allocating more avoidance space to high-risk areas.

[0007] Furthermore, the process of adaptively adjusting cluster size based on environmental complexity, balancing computational load and communication efficiency, and integrating local real-time decision-making with global coordination includes the following steps: Quantitative assessment of environmental complexity: The environmental complexity C is quantified by collecting the following data from edge nodes: Obstacle density ρ: The number of dynamic obstacles per unit area (detected jointly by LiDAR and camera); Dynamic obstacle ratio D: The proportion of dynamic obstacles, such as moving vehicles and pedestrians, to the total number of obstacles; Path intersection frequency F: number of potential intersection points between robot path and dynamic obstacle trajectory (predicted by MCTS); Overall complexity calculation: weighted average Or fuzzy logic fusion of the above indicators, generate a complexity score of 0-100 points; Dynamic adjustment of cluster size: According to the complexity score C, dynamically divide or merge the cluster: High complexity scenario (C>70), dense obstacles or high dynamicity, cluster size is reduced, such as 100m 2 The area is divided into 4 clusters of 25m 2 , reduce the computational load within a single cluster (path planning time is reduced), but increase the communication frequency between clusters; low complexity scenario (C<30), open environment or static obstacles, cluster size is expanded, such as merging adjacent clusters, reducing communication overhead, reducing the number of inter-cluster message passing, but allowing a single cluster to undertake more computing (because of fewer obstacles, low computing pressure); Set complexity threshold (such as ), when C exceeds the threshold, trigger splitting or merging operation, and broadcast cluster boundary update information through edge nodes; Balance computing load and communication efficiency: In small-scale clusters, edge nodes can generate candidate paths in parallel, such as multi-threaded execution of improved A* algorithm, shortening single-cluster decision time (from 500ms of centralized control to within 200ms); high-priority tasks (such as emergency material transportation) are preferentially processed in local clusters to avoid cross-cluster communication delay; Only when the environmental complexity changes (such as C fluctuation exceeds 10%) or cross-cluster conflicts are detected, trigger inter-cluster communication, such as broadcasting path intent or requesting coordination, reduce redundant messages; inter-cluster shared information (such as local map, path intent) uses lightweight protocol to compress data volume, reduce bandwidth occupation; Integrate local real-time decision and global coordination: Based on the cluster environment information, dynamic obstacle trajectory, and other robot path intent, edge nodes generate candidate paths through improved A* algorithm, and select the Pareto optimal path set (considering path length, safety, energy efficiency); for high-risk obstacles detected within the cluster, such as the risk value greater than 80 points in step one, immediately trigger local path re-planning, adjust the robot speed or direction; Edge nodes of adjacent clusters share path intent, predict cross-cluster conflict probability (such as The improved auction algorithm is adopted, and the robots in the cluster generate bids according to the path cost (remaining distance, energy consumption) and task priority. High-priority tasks (such as medical material transportation) obtain higher bidding weights, so that critical paths are preferentially executed. A local optimal path is generated, and conflict areas requiring global coordination, such as cross-cluster path intersections, are marked. Through inter-cluster negotiation, a path adjustment protocol such as detouring or slowing down of certain robots is reached, the global path map is updated, and feedback is provided to the multi-objective optimization module in step three to adjust the path weight to adapt to the coordination result.

[0008] Further, the weights of path length, energy consumption, and safety based on the environmental risk value are dynamically adjusted, including the following steps: Quantization and input of environmental risk value: The dynamic risk assessment result of step one is inherited, that is, the risk value of each dynamic obstacle ( ) and the comprehensive environmental risk value of the current area of the robot ( ); through weighted average or maximum value method, the risk values of multiple obstacles are fused into regional risk indicators, for example: , reflecting the potential danger degree of the current position of the robot; The environmental risk value is mapped to the weights of the three targets by using a Sigmoid function ( , ), so that the weights change nonlinearly with the risk value:

[0009] where k is the sensitivity coefficient, is the adjustment threshold of each target (for example, is the path length weight, is lower; is the safety weight, is higher); The sum of the weights is ensured to be 1 by using the softmax function ( ), avoiding the situation that a single target weight is too high and other targets are ignored; Different weight allocation strategies under different risk values: High-risk scenario ( >80): The safety weight ( ) is increased to 0.6-0.8, and the path away from the high-risk obstacle is preferentially selected, even if the path length or energy consumption increases; the path length and energy weight ( , ) is reduced to 0.2-0.4, allowing detour to avoid risks; Medium-risk scenario (50 <80): The safety weight is maintained at 0.5, while path length and energy each account for 0.25, seeking a compromise between safety and efficiency; Low-risk scenarios <50): The path length weight is increased to 0.5-0.7, prioritizing the shortest path to improve efficiency; the energy efficiency weight is increased to 0.3-0.5, optimizing speed or reducing sharp turns to reduce energy consumption; the safety weight is reduced to 0.2, maintaining only basic obstacle avoidance functions. Based on the adjusted weights, a Pareto front path set is generated using an improved multi-objective gray wolf optimizer, with each path labeled with quantitative scores for path length, energy consumption, and safety.

[0010] Optimal path selection: Based on the current risk value and weight, select the path that meets the dominant objective from the Pareto front. For example, select the path with the highest safety score when the risk is high, and select the path with the shortest length when the risk is low. Transmit the coordinates, speed, avoidance strategy and other information of the selected path to the conflict detection module in step four, and update the global path map. Typical scenario example: Sudden obstacles (high risk): With a risk value of 90, a safety weight of 0.8, and a path length weight of 0.1, the robot prioritizes detours, successfully avoiding high-speed approaching obstacles even if the path length increases by 30%. Open space (low risk): With a risk value of 20, a path length weight of 0.7, and an energy weight of 0.3, the robot selects the shortest path and travels at a constant speed, reducing energy consumption by 15%.

[0011] Furthermore, the prediction of future conflict probabilities based on the robot's current path and dynamic obstacle trajectories using Monte Carlo tree search includes the following steps: Input the robot's current path and the trajectory of dynamic obstacles, wherein the robot's current path is the optimal path coordinate sequence selected in step three (e.g., ...). ={( , ), ( , The dynamic obstacle trajectory is the predicted position coordinates of the obstacle within the next 1-2 seconds, as shown in step one. For example:

[0012] Starting from the beginning of the robot's current path, a sub-path is selected for expansion based on historical simulation experience. At the end of the sub-path, multiple possible future path branches are generated, such as uniform straight-line movement, turning and obstacle avoidance, covering the robot's possible decision options. For each path branch, a scenario where the obstacle moves along the predicted trajectory is simulated, and the number of overlaps between the path branch and the obstacle's trajectory is calculated. and total number of simulations Update the conflict probability of the path branches based on the simulation results. It records path areas with a high probability of conflict, such as path turning points and areas with dense obstacles; generates a conflict probability heatmap, marking the location and probability of potential conflicts within the next 1-2 steps of the robot, such as "path turning point conflict probability 85%", to provide a basis for distributed negotiation.

[0013] Furthermore, the distributed auction mechanism includes the following steps: When MCTS predicts the probability of conflict If the threshold is exceeded, the distributed auction mechanism is activated; The robot determines the remaining distance along the current path. Energy consumption Generate basic bid , where α and β are adjustable coefficients; High-priority tasks, such as medical supply transportation and fire rescue, receive additional bidding bonuses. = ×(1+γU), where, The adjusted bid for high-priority tasks (the final bid value in the auction), γ is the task priority adjustment coefficient, used to control the increase in bid due to priority. For example, γ=0.5 means that the bid can be increased by up to 50% due to priority; U is the task urgency (such as a normalized value of 0-1). Robot's Remaining Energy Below the threshold (e.g.) When the percentage is 30%, the bidding price decreases: ,in, The final bidding value for the robot (adjusted after considering remaining energy) is δ, which is the energy adjustment coefficient used to control the reduction in bidding price due to low power. For example, δ=0.8 means that the bidding price decreases by 0.8% for every 1% decrease in remaining energy. This is to avoid robots with low power taking on high-cost paths. Auction process: Broadcast bidding: All robots within the conflict area broadcast their own bids through edge nodes. And the intended route, such as "planned detour via the left side"; Winner determination: the robot with the highest bid wins the path priority, and other robots need to adjust their paths, such as slowing down or detouring to avoid conflicts; Negotiation feedback: the winning robot broadcasts the adjusted path information to the conflict area, and other robots update the local map and trigger the weight adjustment in step three, such as increasing the safety weight; Predicting conflicts 1-2 steps in advance through MCTS to avoid the delay of real-time negotiation (traditional methods need to negotiate after the conflict occurs, resulting in a 0.5-1 second decision lag); combining task priority and robot state (such as remaining energy) to ensure that critical tasks are executed first, while avoiding low-energy robots from being trapped in dangerous paths due to low bidding; Auction mechanism only in conflict areas, reducing global communication overhead, while ensuring local decision consistency with global path through edge node coordination; For example: Two robot path intersection (high risk): MCTS predicts a 85% conflict probability, triggering the auction mechanism; Robot A (transporting medical supplies, high priority, bid =120) wins, robot B (patrol task, bid =80) adjusts the path to detour right; Low-energy robot encounters dynamic obstacles (medium risk): MCTS predicts a 65% conflict probability, robot C (20% remaining energy, bid =50) bids lower than robot D (normal state, bid =90), robot C slows down to avoid, while triggering local re-planning in step five.

[0014] Further, the historical dynamic scene data is stored in the experience replay pool, and the meta-learning model is trained; when encountering a new scene, the model quickly generates an adaptive strategy through a small amount of gradient updates, including the following steps: Constructing a storage and experience replay pool for historical dynamic scene data: Data sources: Dynamic obstacle trajectory: the future position sequence of the obstacle predicted in step one, such as "obstacle A is located at (2,3) at t=1s and (4,5) at t=2s"; Conflict resolution case: the path adjustment record of successful negotiation through the auction mechanism in step four, such as "robot B chooses to detour right in the conflict area to avoid collision with robot A"; Path adjustment log: local re-planning record triggered by environmental changes in step five, such as "robot C deviates 0.5 meters to the left from the original path due to the sudden appearance of an obstacle"; Environmental risk value and weight adjustment record: the weight configuration corresponding to different risk values in step three, such as "risk value 85, safety weight 0.8, path length weight 0.1"; The above data is classified according to scene types, such as "narrow passage obstacle avoidance" and "multi-robot crossing path", and key indicators are labeled, such as obstacle speed, robot remaining energy, and conflict resolution time. Through time stamp association, multiple source data in the same scene, such as obstacle trajectory, robot decision, and environmental risk value, form a complete scene-decision-result chain; The experience replay pool serves as the "experience library" of the meta-learning model, storing diverse dynamic scene cases to provide materials for the model's generalization learning. Through random sampling or priority sampling, such as prioritizing high-conflict scenes, the model's learning efficiency for key scenes is improved. Training of meta-learning model strategy generalization: The model-independent meta-learning framework is adopted, and its core idea is to train an initial model parameter so that the parameter can quickly adapt to new scenes with only a small amount of gradient update; A batch of scenes are randomly extracted from the experience replay pool, such as "narrow passage obstacle avoidance" and "sudden obstacle detour". Each scene contains input (environmental state, such as obstacle trajectory and robot position) and output (optimal path strategy). For each scene, the initial model parameter is used for forward inference, and the loss function, such as path safety loss and energy efficiency loss, is calculated. Through gradient descent, the model parameter is updated to minimize the comprehensive loss of the model in multiple scenes, forming a "universal" initial parameter. The model's rapid adaptation capability in unseen scenes, such as "dynamic obstacle sudden change of direction", is verified to ensure that a small amount of gradient update (such as 1-5 steps) can generate an effective strategy. Through the embedding layer, scene data (such as obstacle trajectory and risk value) is mapped into a low-dimensional feature vector, improving the model's perception of scene differences. Combined with multi-objective optimization indicators (path length, safety, and energy efficiency), a differentiable loss function is designed to support end-to-end training. Rapid adaptation strategy generation in new scenes: When the robot encounters a new scene, such as obstacle motion pattern difference exceeding the threshold compared to historical data, scene features are extracted through the environmental perception module in step one, such as obstacle speed distribution and path intersection frequency. Based on the MAML framework, a small amount of samples (such as 1-5 time steps of obstacle trajectory) of the new scene are used to calculate the gradient and update the model parameter. The updated parameter retains the generalization ability for general scenes while fine-tuning specific features (such as high-speed obstacle change of direction) of the new scene to generate an adaptation strategy (such as "start detour 0.5 seconds in advance"). Generate path planning strategies that adapt to new scenarios, such as detour directions and speed adjustment ranges, and label the confidence of the strategies, such as "left detour strategy confidence 90%"; if the confidence is below the threshold, trigger the distributed auction mechanism in step four to further optimize the strategy through negotiation; Typical scenario examples: Sudden obstacle change direction (new scenario): The meta-learning model adjusts the parameters using the 3-time step data of the new scenario (the positions before and after the obstacle turns) through the meta-update step, generating a "slow down and detour to the right" strategy to avoid collision. Multi-robot dense collaboration (general scenario): The model directly outputs the strategy "distributed auction mechanism triggered, high-priority robot detours first" based on the "multi-robot crossing path" cases in the experience replay pool, without the need for additional fine-tuning.

[0015] A path planning system for multiple mobile robots, comprising: Environment perception modeling layer: Real-time collection of environmental data through multi-sensor fusion (laser radar, camera, IMU), construction of high-precision map; prediction of dynamic obstacle future trajectory using lightweight LSTM network deployed on edge node, and calculation of risk value combining obstacle speed, direction, and size; finally output dynamic environment model labeled with risk level, providing real-time and accurate perception basis for subsequent path planning; Distributed cooperative path planning layer: dynamically divide clusters according to robot position, task priority, and communication range, and manage intra-cluster cooperation by edge nodes; dynamically adjust the weights of path length, energy consumption, and safety based on environmental risk value, generate Pareto optimal path set through improved multi-objective grey wolf optimizer; combine improved A* algorithm to generate local candidate paths, and select the optimal path that meets multi-objective constraints for local real-time decision and global coordination fusion; Dynamic coordination conflict resolution layer: based on the current path of the robot and the trajectory of the dynamic obstacle, predict the conflict probability of the next 1-2 steps through Monte Carlo tree search; when the conflict probability exceeds the threshold, trigger the distributed auction mechanism, and robots generate bids based on path cost, task priority, and remaining energy; high-priority tasks get higher weights; adjust the path through negotiation to avoid conflicts, and feed the results back to the path planning layer to form a closed-loop coordination; Continuous learning optimization layer: store historical dynamic scenario data in the experience replay pool and train the meta-learning model; when encountering a new scenario, the model quickly adapts through a small amount of gradient updates to generate an adaptive strategy; the optimized strategy and weight adjustment rules are injected back into the previous modules to form an iterative optimization closed loop of "perception-planning-coordination-learning", improving the system's generalization ability to unknown scenarios.

[0016] The present application has the following advantages: In the present application, multi-sensor fusion and edge node LSTM prediction are used to provide real-time and accurate dynamic environment information; through dynamic clustering and edge node management, local real-time decision and global coordination are realized, and the calculation load is balanced; and based on the environmental risk value, the weights of path length, energy consumption and safety are dynamically adjusted, and the improved grey wolf optimizer is combined to balance multiple objectives in real time; through Monte Carlo tree search, the conflict probability is predicted 1-2 steps in advance, the distributed auction mechanism is triggered, the bidding is dynamically adjusted combined with the task priority and the remaining energy, and the conflict is quickly negotiated and solved; and the meta-learning model is used to learn from historical scenes, and when a new scene is encountered, a small amount of gradient update is used to quickly generate an adaptive strategy, realizing "zero sample" or "small sample" scene adaptation, effectively solving the problem of insufficient dynamic environment adaptability of the existing path planning method of multiple mobile robots. BRIEF DESCRIPTION OF DRAWINGS

[0017] Fig. 1 The steps block diagram of the path planning method of the multiple mobile robots proposed in the present application; Fig. 2 The system block diagram of the path planning system of the multiple mobile robots proposed in the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] Please refer to Figs. 1-2 The present application is a path planning method and system for multiple mobile robots, which includes: The path planning method for multiple mobile robots includes: Step 1: Real-time construction of environment map through multi-sensor fusion (laser radar, camera, IMU), labeling of dynamic obstacle (such as moving target, sudden obstacle) motion trajectory; using the lightweight neural network deployed by the edge node to quickly identify the obstacle type (static / dynamic), and predicting its future position through LSTM; combining the dynamic obstacle risk level evaluation model, calculating the collision risk value based on speed, direction and size, and processing high-risk targets preferentially, to provide real-time and accurate dynamic environment information for subsequent path planning; Step two: dynamically divide clusters according to robot positions, task priorities, and communication ranges, with each cluster managed by an edge node (such as a roadside unit); robots within a cluster share local maps and path intentions, and the edge node generates candidate paths using an improved A* algorithm, and filters out the Pareto optimal path set based on non-dominated sorting and crowding comparison; adjust the cluster size adaptively based on environmental complexity (such as obstacle density) to balance computational load and communication efficiency, and integrate local real-time decision-making and global coordination; Step three: dynamically adjust the weights of path length, energy consumption, and safety based on environmental risk values to avoid suboptimal solutions caused by fixed weights; introduce an improved multi-objective grey wolf optimizer, combine non-dominated sorting and crowding comparison to generate Pareto frontier paths; use a weight adaptive mechanism based on environmental perception to balance multiple objectives in real time, such as prioritizing shortening paths for urgent tasks and prioritizing energy-saving paths for low-energy tasks, to improve comprehensive optimization capabilities in dynamic scenarios; Step four: based on the current path of the robot and the trajectory of dynamic obstacles, predict the probability of future conflicts through Monte Carlo tree search; use a distributed auction mechanism, and robots generate bids based on path cost (remaining distance, energy consumption) and task priority, with high-priority tasks (such as medical supply transportation) receiving higher bidding weights; dynamically adjust the bidding weight based on task urgency and robot remaining energy to predict conflicts 1-2 steps in advance and negotiate solutions, reducing the number of real-time negotiations, communication delays, and conflict rates; Step five: robots use odometry and visual odometry to correct path deviations in real time, and trigger local re-planning if the deviation exceeds a threshold; store historical dynamic scenario data (obstacle movement patterns, conflict resolution cases) in an experience replay pool to train a meta-learning model; when encountering a new scenario, the model quickly generates an adaptive strategy through a small number of gradient updates, enabling "zero-shot" or "few-shot" scenario adaptation, breaking the traditional deep learning dependence on training data, and improving generalization ability and real-time performance in dynamic environments.

[0020] In one embodiment, the lightweight neural network deployed by the edge node quickly identifies the type of obstacle and predicts its future position through LSTM; combined with a dynamic obstacle risk level evaluation model, high-risk targets are prioritized, including the following steps: Obstacle type identification using lightweight neural network: Edge nodes (such as robot local or roadside units) receive multi-sensor fusion data (laser radar point cloud, camera image, IMU attitude information), forming a raw data stream containing obstacle geometric features and motion state; Deploy improved MobileNet or ShuffleNet lightweight neural network, extract obstacle visual features (such as shape, texture) and motion features (such as displacement between consecutive frames) through convolutional layers, and finally determine obstacle type (static / dynamic) through classification layer; Network model through quantization, pruning optimization means, make the inference speed on embedded device reach 10-20ms level, meet the real-time identification demand in dynamic scene; Trajectory prediction of LSTM network: Take the position coordinates of dynamic obstacles in the past 3-5 seconds, such as laser radar detection two-dimensional / three-dimensional points, as input sequence to construct time window data; Through the hidden state memory mechanism of LSTM unit, learn the motion mode of obstacles, such as uniform straight line, acceleration, turning, etc., output the predicted trajectory in the future 1-2 seconds in the form of probability distribution, such as Gaussian process or particle filter result; Score the confidence of the predicted trajectory, if the confidence is lower than the threshold, trigger sensor resampling or fuse other sensor data such as radar speed information to improve the prediction reliability; Dynamic risk level evaluation: Risk index calculation: Speed risk: the radial speed of the obstacle relative to the robot Compared with the threshold The ratio =+ , the risk of high-speed approaching obstacle is higher; Direction risk: the cosine value of the angle θ between the motion direction of the obstacle and the robot path , the risk of obstacle towards the robot path is higher; Size risk: the ratio of the volume V occupied by the obstacle to the safety threshold The risk of large obstacle is higher; Through weighted summation Or fuzzy logic fusion of the above indexes, generate a risk value of 0-100 points; According to the risk value, for the obstacle with risk value greater than 80 points (such as large vehicle approaching at high speed), trigger local path re-planning immediately, and adjust the robot motion speed, such as deceleration or detour; For the obstacle with risk value less than 50 points, such as slow-moving pedestrians, only keep monitoring, do not interfere with path planning; ​The identified obstacle types, predicted trajectories, risk values, and priority labels are updated in real time to the global environment map and broadcast to other robots within the cluster (through edge nodes); information of high-risk targets is directly input into the dynamic clustering module in step two, affecting the cluster's path generation and coordination strategy, such as allocating more avoidance space for high-risk areas.

[0021] In one embodiment, the adaptive adjustment of cluster size based on environmental complexity balances the computational load and communication efficiency, and integrates local real-time decision-making and global coordination, including the following steps: Quantitative assessment of environmental complexity: Through the edge node, the following data is collected to quantify the environmental complexity C: Obstacle density p: the number of dynamic obstacles per unit area (detected by laser radar and camera combined); Dynamic obstacle proportion D: the proportion of dynamic obstacles, such as moving vehicles and pedestrians, in the total obstacles; Path intersection frequency F: the number of potential intersection points between robot paths and dynamic obstacle trajectories (predicted by MCTS); Comprehensive complexity calculation: weighted average Or fuzzy logic fusion of the above indicators to generate a complexity score of 0-100; Dynamic adjustment of cluster size: Dynamic division or merging of clusters based on complexity score C: High complexity scenario (C>70), obstacle density or dynamicity is high, cluster size is reduced, such as dividing a 100m 2 area into 4 clusters of 25m 2 each, reducing the computational load within a single cluster (path planning time is reduced), but increasing the communication frequency between clusters; low complexity scenario (C<30), open environment or static obstacles are the main, cluster size is expanded, such as merging adjacent clusters, reducing communication overhead, the number of message passing between clusters is reduced, but allowing a single cluster to bear more computational load (due to fewer obstacles, low computational pressure); Set a complexity threshold (e.g. ), when C exceeds the threshold, trigger splitting or merging operation, and broadcast cluster boundary update information through edge nodes; Balancing computational load and communication efficiency: In small-scale clusters, edge nodes can generate candidate paths in parallel, such as multi-threaded execution of improved A* algorithm, shortening the single-cluster decision-making time (from 500ms in centralized control to within 200ms); high-priority tasks (such as emergency material transportation) are preferentially processed in local clusters to avoid cross-cluster communication delay; Inter-cluster communication is triggered only when environmental complexity changes (e.g. C fluctuation exceeds 10%) or cross-cluster conflict is detected, such as broadcasting path intention or requesting coordination, reducing redundant messages; inter-cluster sharing of information (e.g. local map, path intention) uses lightweight protocol to compress data volume and reduce bandwidth occupancy; Fusion of local real-time decision and global coordination: Based on intra-cluster environmental information, dynamic obstacle trajectory, and other robot path intention, the edge node generates candidate paths by improving the A* algorithm and selects the Pareto optimal path set (considering path length, safety, and energy efficiency); for high-risk obstacles detected within the cluster, such as the risk value greater than 80 points evaluated in step one, local path re-planning is triggered immediately to adjust the robot speed or direction; The edge nodes of adjacent clusters share path intentions and predict cross-cluster conflict probability through geometric calculation (e.g. path bounding box overlap detection) or MCTS simulation ( >τ triggers coordination); an improved auction algorithm is used, and inter-cluster robots generate bids based on path cost (remaining distance, energy consumption) and task priority, with high-priority tasks (such as medical supply transportation) obtaining higher bidding weights to prioritize critical paths; Generate local optimal paths and label conflict areas requiring global coordination, such as cross-cluster path intersections; reach path adjustment agreements through inter-cluster negotiation, such as certain robots taking detours or reducing speed, update the global path map, and feedback to the multi-objective optimization module in step three to adjust path weights to adapt to coordination results.

[0022] In one embodiment, the dynamic adjustment of path length, energy consumption, and safety weight based on environmental risk value includes the following steps: Quantification and input of environmental risk value: Inherit the dynamic risk assessment results of step one, i.e. the risk value of each dynamic obstacle ( ), and the comprehensive environmental risk value of the area where the robot is currently located ( ); by weighted average or maximum value method, the risk values of multiple obstacles are fused into regional risk indicators, for example: , reflecting the potential danger level of the robot's current position; Use a Sigmoid function to map the environmental risk value to the weights of the three objectives ( , ), making the weight change nonlinearly with the risk value:

[0023] where k is the sensitivity coefficient, is the adjustment threshold of each objective (e.g. is the path length weight, Lower; For security weights, (Higher) The softmax function ensures that the sum of the weights is 1. This avoids giving too much weight to a single objective, causing other objectives to be overlooked. Weighting strategies for different risk values: High-risk scenarios >80): Security weights ( The energy weight is increased to 0.6-0.8, prioritizing paths away from high-risk obstacles, even if path length or energy consumption increases; path length and energy weight ( , The risk level should be reduced to 0.2-0.4, allowing detours to mitigate the risk. Medium-risk scenarios (50<) <80): The safety weight is maintained at 0.5, while path length and energy each account for 0.25, seeking a compromise between safety and efficiency; Low-risk scenarios <50): The path length weight is increased to 0.5-0.7, prioritizing the shortest path to improve efficiency; the energy efficiency weight is increased to 0.3-0.5, optimizing speed or reducing sharp turns to reduce energy consumption; the safety weight is reduced to 0.2, maintaining only basic obstacle avoidance functions. Based on the adjusted weights, a Pareto front path set is generated using an improved multi-objective gray wolf optimizer, with each path labeled with quantitative scores for path length, energy consumption, and safety.

[0024] Optimal path selection: Based on the current risk value and weight, select the path that meets the dominant objective from the Pareto front. For example, select the path with the highest safety score when the risk is high, and select the path with the shortest length when the risk is low. Transmit the coordinates, speed, avoidance strategy and other information of the selected path to the conflict detection module in step four, and update the global path map. Typical scenario example: Sudden obstacles (high risk): With a risk value of 90, a safety weight of 0.8, and a path length weight of 0.1, the robot prioritizes detours, successfully avoiding high-speed approaching obstacles even if the path length increases by 30%. Open space (low risk): With a risk value of 20, a path length weight of 0.7, and an energy weight of 0.3, the robot selects the shortest path and travels at a constant speed, reducing energy consumption by 15%.

[0025] In one embodiment, predicting the probability of future conflicts based on the robot's current path and the trajectory of dynamic obstacles using Monte Carlo tree search includes the following steps: Input the robot's current path and the trajectory of dynamic obstacles, wherein the robot's current path is the optimal path coordinate sequence selected in step three (e.g., ...). ={( , ), ( , The dynamic obstacle trajectory is the predicted position coordinates of the obstacle within the next 1-2 seconds, as shown in step one. For example:

[0026] Starting from the beginning of the robot's current path, a sub-path is selected for expansion based on historical simulation experience. At the end of the sub-path, multiple possible future path branches are generated, such as uniform straight-line movement, turning and obstacle avoidance, covering the robot's possible decision options. For each path branch, a scenario where the obstacle moves along the predicted trajectory is simulated, and the number of overlaps between the path branch and the obstacle's trajectory is calculated. and total number of simulations Update the conflict probability of the path branches based on the simulation results. It records path areas with a high probability of conflict, such as path turning points and areas with dense obstacles; generates a conflict probability heatmap, marking the location and probability of potential conflicts within the next 1-2 steps of the robot, such as "path turning point conflict probability 85%", to provide a basis for distributed negotiation.

[0027] Furthermore, the distributed auction mechanism includes the following steps: When MCTS predicts the probability of conflict If the threshold is exceeded, the distributed auction mechanism is activated; The robot determines the remaining distance along the current path. Energy consumption Generate basic bid , where α and β are adjustable coefficients; High-priority tasks, such as medical supply transportation and fire rescue, receive additional bidding bonuses. = ×(1+γU), where, The adjusted bid for high-priority tasks (the final bid value in the auction), γ is the task priority adjustment coefficient, used to control the increase in bid due to priority. For example, γ=0.5 means that the bid can be increased by up to 50% due to priority; U is the task urgency (such as a normalized value of 0-1). Robot's Remaining Energy Below the threshold (e.g.) When the percentage is 30%, the bidding price decreases: wherein, is the final bid value of the robot for the auction (considering the adjustment result after the remaining energy), and δ is an energy adjustment coefficient for controlling the reduction of the bid value due to low energy, for example, δ = 0.8 means that the bid is reduced by 0.8% for every 1% reduction in the remaining energy; and is to avoid a low-energy robot taking a high-cost path; Auction process: Broadcast bid: all robots in the conflict area broadcast their own bid and path intention through the edge node, such as "plan to detour left"; The winner is determined: the robot with the highest bid value obtains the path priority, and other robots need to adjust their paths, such as slowing down or detouring to avoid conflicts; Negotiation feedback: the winning robot broadcasts the adjusted path information to the conflict area, and other robots update the local map and trigger the weight adjustment in step three, such as increasing the safety weight; Predict conflicts 1-2 steps in advance through MCTS to avoid the delay of real-time negotiation (traditional methods need to negotiate after the conflict occurs, resulting in a 0.5-1 second decision lag); combine task priority and robot state (such as remaining energy) to ensure that critical tasks are executed first, while avoiding low-energy robots from getting into dangerous paths due to low bids; the auction mechanism is only executed in the conflict area, reducing the global communication overhead, while the edge node coordinates to ensure the consistency of local decisions and global paths; for example: Two robot paths intersect (high risk): MCTS predicts a conflict probability of 85%, triggering the auction mechanism; Robot A (transporting medical supplies, high priority, bid =120) wins, and robot B (patrol task, bid =80) adjusts the path to detour right; Robot low energy encounters dynamic obstacles (medium risk): MCTS predicts a conflict probability of 65%, and robot C (remaining energy 20%, bid =50) has a lower bid than robot D (normal state, bid =90), and robot C slows down to avoid, while triggering local re-planning in step five. In one embodiment, the historical dynamic scene data is stored in an experience replay pool, and a meta-learning model is trained; when a new scene is encountered, the model quickly generates an adaptive strategy through a small number of gradient updates, including the following steps:

[0028] Construct a storage and experience replay pool for historical dynamic scene data: Data sources: ​Dynamic obstacle trajectory: The sequence of future obstacle positions predicted in step one, such as "obstacle A is located at (2,3) at t=1s and at (4,5) at t=2s"; Conflict resolution case: The path adjustment record successfully negotiated through the auction mechanism in step four, such as "Robot B chooses to detour to the right in the conflict area to avoid colliding with Robot A"; Path adjustment log: Records of local replanning triggered by environmental changes in step five, such as "Robot C shifted 0.5 meters to the left from its original path due to the sudden appearance of an obstacle"; Environmental risk value and weight adjustment record: The weight configuration corresponding to different risk values ​​in step three, such as "when the risk value is 85, the safety weight is 0.8 and the path length weight is 0.1"; The above data is categorized by scenario type, such as "obstacle avoidance in narrow passages" and "multi-robot cross-path", and key indicators are labeled, such as obstacle speed, robot remaining energy, and conflict resolution time. Multi-source data in the same scenario, such as obstacle trajectory, robot decision, and environmental risk value, are linked by timestamps to form a complete scenario-decision-result chain. The experience replay pool serves as the "experience repository" for the meta-learning model, storing diverse dynamic scene cases to provide materials for the model's generalization learning. Through random sampling or priority sampling, such as prioritizing high-conflict scenes, the model's learning efficiency for key scenes is improved. Training meta-learning model policy generalization: The core idea of ​​the model-independent meta-learning framework is to train an initial model with parameters that can quickly adapt to new scenarios with only a small number of gradient updates. A batch of scenarios, such as "obstacle avoidance in narrow passages" and "obstacle detour," are randomly selected from the experience replay pool. Each scenario includes input (environmental state, such as obstacle trajectory and robot position) and output (optimal path strategy). For each scenario, forward inference is performed using initial model parameters to calculate loss functions, such as path safety loss and energy efficiency loss. The model parameters are updated through gradient descent to minimize the overall loss of the model across multiple scenarios, forming "general" initial parameters. Validate the model’s ability to quickly adapt to unseen scenarios, such as “sudden change of direction by dynamic obstacles”, and ensure that an effective policy can be generated with a small number of gradient updates (e.g., 1-5 steps). By mapping scene data (such as obstacle trajectories and risk values) into low-dimensional feature vectors through an embedding layer, the model's ability to perceive scene differences is improved; combined with multi-objective optimization metrics (path length, safety, and energy efficiency), a differentiable loss function is designed to support end-to-end training. Generate a fast adaptation strategy for new scenarios: When the robot encounters a new scene, such as the obstacle motion pattern differs from the historical data by more than a threshold, the environment perception module in step one extracts scene features such as obstacle speed distribution, path intersection frequency; Based on the MAML framework, use a small number of samples (such as 1-5 time steps of obstacle trajectory) of the new scene to calculate the gradient and update the model parameters; the updated parameters retain the generalization ability to the general scene, while fine-tuning the specific features of the new scene (such as high-speed obstacle turning), generating an adaptive strategy (such as "start turning 0.5 seconds in advance"); Generate path planning strategies that adapt to new scenes such as turning direction, speed adjustment amplitude, and label the confidence of the strategy, such as "turning left strategy confidence 90%"; if the confidence is below a threshold, trigger the distributed auction mechanism in step four to further optimize the strategy; for example: Sudden obstacle turning (new scene): The meta-learning model generates a "decelerate in advance and turn right" strategy to avoid collision by adjusting the parameters using 3 time step data (obstacle turning position before and after) of the new scene through the meta-update step, as the historical data does not contain the case of "obstacle turning 90 degrees suddenly"; Multi-robot dense collaboration (general scene): The model directly outputs the strategy "distributed auction mechanism triggered, high-priority robot turns first" based on the "multi-robot crossing path" case in the experience replay pool, without the need for additional fine-tuning.

[0029] A path planning system for multiple mobile robots, comprising: Environment perception modeling layer: real-time collection of environmental data through multi-sensor fusion (laser radar, camera, IMU), construction of high-precision map; prediction of future trajectory of dynamic obstacles using lightweight LSTM network deployed on edge nodes, and calculation of risk value combining obstacle speed, direction, and size; finally output a dynamic environment model labeled with risk level, providing real-time and accurate perception basis for subsequent path planning; Distributed cooperative path planning layer: dynamically divide clusters according to robot position, task priority, and communication range, managed by edge nodes within the cluster; dynamically adjust the weights of path length, energy consumption, and safety based on environmental risk value, generate Pareto optimal path set through improved multi-objective grey wolf optimizer; combine improved A* algorithm to generate local candidate paths, and select the optimal path that meets the multi-objective constraints for local real-time decision and global coordination fusion; Dynamic coordination conflict resolution layer: based on the current path of the robot and the trajectory of the dynamic obstacle, the conflict probability of the next 1-2 steps is predicted through Monte Carlo tree search; when the conflict probability exceeds the threshold, the distributed auction mechanism is triggered, and the robot generates the bid according to the path cost, task priority and remaining energy, and the high priority task obtains higher weight; the path is adjusted through negotiation to avoid conflict, and the result is fed back to the path planning layer to form a closed loop coordination; Continuous learning optimization layer: store historical dynamic scene data into experience replay pool, train meta-learning model; when encountering new scene, the model quickly adapts through a small amount of gradient update to generate adaptive strategy; the optimized strategy and weight adjustment rule are injected into the previous module in reverse, forming an iterative optimization closed loop of "perception-planning-coordination-learning", improving the generalization ability of the system to unknown scenes.

[0030] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for path planning of multiple mobile robots, characterized by, Comprise: Step one: Real-time construction of environment map and labeling of dynamic obstacle trajectory through multi-sensor fusion; Lightweight neural network deployed on edge node is used to quickly identify obstacle type and predict future position through LSTM; combined with dynamic obstacle risk level evaluation model, high-risk targets are preferentially processed, providing real-time dynamic environment information for path planning; Step two: Based on robot position, task priority, and communication range, clusters are dynamically divided, and each cluster is managed by an edge node; robots in the cluster share local map and path intention, and the edge node predicts the future 1-2 second trajectory of dynamic obstacles through LSTM, and combines the environmental risk value to nonlinearly map the path length, energy consumption, and safety weight to generate candidate paths in dynamic environment using A* algorithm; based on non-dominated sorting and congestion comparison, the Pareto optimal path set is selected; the cluster size is adaptively adjusted based on environmental complexity to balance the calculation load and communication efficiency, and the fusion of local real-time decision and global coordination is performed; Step three: Based on environmental risk value, the weights of path length, energy consumption, and safety are dynamically adjusted; A multi-objective grey wolf optimizer with dynamic weight adjustment and multi-objective optimization mechanism is used to generate Pareto frontier paths through non-dominated sorting and congestion comparison; real-time multi-objective balance is achieved through weight adaptive mechanism based on environmental perception; Step four: Based on the current path of the robot and the trajectory of the dynamic obstacle, the future conflict probability is predicted through Monte Carlo tree search; A distributed auction mechanism is used, and robots generate bids based on path cost and task priority; high-priority tasks have higher bidding weights; task urgency and robot remaining energy are used to dynamically adjust the bidding weight, predict conflicts 1-2 steps in advance, and negotiate solutions; Step five: The robot corrects the path deviation in real time through odometry and visual odometry, and triggers local replanning if the deviation exceeds the threshold; historical dynamic scene data is stored in the experience replay pool to train the meta-learning model; when encountering a new scene, the model quickly generates an adaptive strategy through a small amount of gradient updates for scene adaptation.

2. The method of claim 1, wherein, The lightweight neural network deployed on the edge node is used to quickly identify the type of obstacle and predict its future position through LSTM; combined with a dynamic obstacle risk level evaluation model, high-risk targets are preferentially processed, including the following steps: The edge node receives multi-sensor fusion data to form an original data stream input lightweight neural network containing obstacle geometric features and motion state; the lightweight neural network refers to compressing the parameters of MobileNetV3 model to 1 / 4 of the original size by 8-bit integerization, using structured pruning to remove redundant convolution kernels and unstructured pruning to remove low-weight connections, and introducing depth separable convolution to decompose standard convolution into depth convolution and point convolution to obtain an improved MobileNetV3 lightweight model; the visual and motion features of the obstacle are extracted through the convolution layer, and finally the obstacle type is determined through the classification layer, achieving a reasoning speed of 10-20 ms on embedded devices; The position coordinates of the dynamic obstacles in the past 3-5 seconds are taken as input sequences to construct time window data; through the hidden state memory mechanism of the LSTM unit, the motion mode of the obstacles is learned, and the predicted trajectory in the future 1-2 seconds is output; the predicted trajectory is scored for confidence, and if the confidence is lower than a threshold, the sensor resampling or fusion of other sensor data is triggered; Risk index calculation: Velocity risk: the radial velocity of the obstacle relative to the robot vs. a threshold =+ Higher velocity = higher risk of a high-speed approach​ Direction risk: the cosine of the angle θ between the direction of the obstacle motion and the robot path , the obstacle is more dangerous towards the robot path; Size risk: the ratio of the volume V of space occupied by the obstacle to the safety threshold Large obstacles are at higher risk.​ by weighted sum or fuzzy logic fusion, generate a risk value between 0 and 100; according to the risk value ranking, immediately trigger local path re-planning and adjust the robot motion speed for obstacles with risk value greater than 80; only keep monitoring for obstacles with risk value less than 50, without intervening in path planning for the moment; The identified obstacle type, predicted trajectory, risk value and priority are marked, and real-time updates are made to the global environment map and broadcast to other robots in the cluster; the information of high-risk targets is directly input into the dynamic clustering module of step two.

3. The method of claim 1, wherein, The cluster size is adaptively adjusted according to the environmental complexity, the calculation load and the communication efficiency are balanced, and the fusion of local real-time decision and global coordination is carried out, including the following steps: The edge nodes statistically quantify the environmental complexity C by the following data: Obstacle density ρ: the number of dynamic obstacles per unit area; Dynamic obstacle ratio D: the proportion of dynamic obstacles in total obstacles; Path intersection frequency F: the number of potential intersection points between the robot path and the dynamic obstacle trajectory; using a weighted average or fuzzy logic fuses the obstacle density p, the dynamic obstacle proportion D, and the path intersection frequency F to generate a complexity score of 0-100. According to the complexity score C, the clusters are dynamically divided or merged: In high-complexity scenarios, obstacles are dense or dynamic, the cluster size is reduced, the calculation load in a single cluster is reduced, but the inter-cluster communication frequency is increased; in low-complexity scenarios, the environment is open or static obstacles are dominant, the cluster size is expanded, the communication overhead is reduced, but a single cluster is allowed to bear a larger calculation load; Set the complexity threshold, trigger the splitting or merging operation when C exceeds the threshold, and broadcast the cluster boundary update information through the edge node; Balance calculation load and communication efficiency: In small-scale clusters, edge nodes can generate candidate paths in parallel to shorten single-cluster decision-making time; high-priority tasks are preferentially processed in local clusters; inter-cluster communication is triggered only when the environmental complexity changes or a cross-cluster conflict is detected, reducing redundant messages; inter-cluster information sharing uses a lightweight protocol to compress data volume and reduce bandwidth occupancy; Fusion of local real-time decision and global coordination: Edge nodes generate candidate paths based on cluster environmental information through the improved A* algorithm; the improved A* algorithm dynamically adjusts the heuristic function, combines the nonlinear mapping of environmental risk values with the weights of path length, energy consumption and safety, integrates dynamic obstacle trajectory prediction, uses LSTM to predict the position of obstacles in the next 1-3 seconds to avoid collisions, and uses non-dominated sorting and congestion comparison to filter the Pareto optimal path set, adapting from static environments to dynamic complex scenarios, while supporting multi-objective real-time balancing, and immediately triggering local path replanning for high-risk obstacles detected in the cluster to adjust the robot speed or direction; Adjacent cluster edge nodes share path intentions, predict cross-cluster conflict probability through geometric calculation or MCTS simulation; using an auction algorithm, inter-cluster robots generate bids according to path cost and task priority, high-priority tasks get higher bidding weights, making key paths preferentially executed; local optimal paths are generated and conflict areas requiring global coordination are marked; a path adjustment protocol is reached through inter-cluster negotiation, the global path map is updated, and feedback is provided to the multi-objective optimization module of step three.

4. The method of claim 1, wherein, The weight of path length, energy consumption and safety is dynamically adjusted based on the environmental risk value, including the following steps: Quantification and input of environmental risk value: Inherit the dynamic risk assessment result of step one, that is, the risk value of each dynamic obstacle , and the comprehensive environmental risk value of the area where the robot is currently located ; the risk values of multiple obstacles are fused into a regional risk index by weighted average or maximum value method: reflecting a degree of potential danger of a current position of the robot; The sigmoid function is used to map the environmental risk value to the weights of the three objectives (Equations 1-3), making the weights vary nonlinearly with the risk value: , , .

5. wherein, k is a sensitivity coefficient, is the adjustment threshold for each target; Ensure the total weight is 1 through the softmax function; Weight distribution strategy under different risk values: High-risk scenario: Safety weight Increased to 0.6-0.8, prefer paths that are further away from high risk obstacles, even if path length or energy consumption increases; path length vs. energy weight , Decreased to 0.2-0.4, allow detours to avoid risk; Medium-risk scenario: Safety weight remains 0.5, path length and energy each account for 0.25; Low-risk scenario: Path length weight increases to 0.5-0.7, preferring the shortest path; energy efficiency weight increases to 0.3-0.5, optimizing speed or reducing sharp turns to reduce energy consumption; safety weight decreases to 0.2, only maintaining basic obstacle avoidance function; Based on the adjusted weight, an improved multi-objective grey wolf optimizer is used to generate a Pareto frontier path set, each path is labeled with the quantified scores of path length, energy consumption and safety; the improved multi-objective grey wolf optimizer generates a Pareto frontier path set through a dynamic weight distribution mechanism combined with non-dominated sorting and crowding comparison, and introduces a grey wolf population position update strategy, upgrading from fixed weight to dynamic multi-objective balance, generating optimal paths that balance path length, energy efficiency and safety in dynamic scenarios in real time; the coordinates, speed and avoidance strategy information of the selected path are transmitted to the conflict detection module in step four, and the global path map is updated.

6. The method of claim 1, wherein, The future conflict probability is predicted based on the current path of the robot and the trajectory of the dynamic obstacle through Monte Carlo tree search, including the following steps: Input the current path of the robot and the trajectory of the dynamic obstacle, where the current path of the robot is the optimal path coordinate sequence selected in step three; the trajectory of the dynamic obstacle is the position coordinates of the obstacle within 1-2 seconds in the future predicted in step one; From the start of the current path of the robot, a sub-path is selected for expansion according to historical simulation experience; at the end of the sub-path, multiple possible future path branches are generated to cover the possible decision options of the robot; for each path branch, the scenario of the obstacle moving according to the predicted trajectory is simulated, and the number of overlaps between the path branch and the obstacle trajectory is calculated and the total number of simulations ; the conflict probability of the path branch is updated according to the simulation result , and the path area with high probability of conflict is recorded; a conflict probability heat map is generated to mark the position and probability of possible conflict within 1-2 steps in the future of the robot.

7. The method of claim 1, wherein, The distributed auction mechanism includes the following steps: When MCTS predicts the probability of conflict If the threshold is exceeded, the distributed auction mechanism is activated; The robot according to the remaining distance of the current path , energy consumption Generating a base bid where a, b are adjustable coefficients; High priority tasks get an additional bid bonus = × (1 + γU), where, is the adjusted bid for a high priority task, γ is the adjustment factor for task priority, and U is the task urgency. Robot residual energy When below threshold, bid is reduced: wherein, is the final participation value of the robot in the auction, and δ is an energy adjustment coefficient used to control the reduction in the bid value due to low energy. Auction process: Broadcast bid: all robots in the conflict area broadcast their bids through the edge nodes and path intent; Winner determination: the robot with the highest bid obtains path priority, and other robots need to adjust their paths to avoid conflicts; Negotiation feedback: the winning robot broadcasts the adjusted path information to the conflict area, and other robots update the local map and trigger the weight adjustment in step three; Predict conflicts 1-2 steps in advance through MCTS to avoid the delay of real-time negotiation; combine task priority and robot state to prioritize key tasks, while avoiding low-bid robots from getting into dangerous paths; the auction mechanism is only executed in the conflict area, reducing global communication overhead, and local decisions are consistent with global paths through edge node coordination.

8. The method of claim 1, wherein, The historical dynamic scenario data is stored in the experience replay pool to train the meta-learning model; when encountering a new scenario, the model quickly generates an adaptive strategy through a small amount of gradient updates, including the following steps: Build a storage and experience replay pool for historical dynamic scenario data: Data sources: Dynamic obstacle trajectory: the future position sequence of the obstacle predicted in step one; Conflict resolution cases: path adjustment records successfully negotiated through the auction mechanism in step four; Path adjustment log: local re-planning records triggered by environmental changes in step five; Environmental risk value and weight adjustment record: weight configuration corresponding to different risk values in step three; The data is classified by scene type and labeled with indicators. The multi-source data under the same scene is associated through timestamps to form a complete scene-decision-result chain. The experience replay pool serves as the "experience base" of the meta-learning model, storing diverse dynamic scene cases to provide materials for the model's generalization learning. Through random sampling or priority sampling, the model's learning efficiency for scenes is improved. Training the meta-learning model strategy generalization: An initial model parameter is trained using a model-independent meta-learning framework, allowing the parameter to quickly adapt to new scenes with only a small amount of gradient updates. A batch of scenes is randomly selected from the experience replay pool, each containing input and output. For each scene, the initial model parameter is used for forward inference, and the loss function is calculated. The model parameter is updated through gradient descent to minimize the comprehensive loss on multiple scenes, forming a "universal" initial parameter. The model's rapid adaptation capability in unseen scenes is verified, allowing a small amount of gradient updates to generate an effective strategy. The scene data is mapped to a low-dimensional feature vector through an embedding layer, improving the model's ability to perceive scene differences. A differentiable loss function is designed in combination with multi-objective optimization indicators to support end-to-end training. Generating a rapid adaptation strategy in a new scene: When the robot encounters a new scene, the scene features are extracted through the environment perception module in step one. Based on the MAML framework, a small number of samples from the new scene are used to calculate the gradient and update the model parameter. The updated parameter retains the generalization ability for universal scenes while fine-tuning specific features of the new scene to generate an adaptation strategy. An adaptation strategy for the new scene is generated, and the confidence of the strategy is labeled. If the confidence is below a threshold, the distributed auction mechanism in step four is triggered to further optimize the strategy through negotiation.

9. The multi-mobile robot path planning system for use in the multi-mobile robot path planning method according to claim 1, characterized by, It includes: Environment perception modeling layer: real-time collection of environmental data through multi-sensor fusion to build a map; A lightweight LSTM network deployed on edge nodes is used to predict the future trajectory of dynamic obstacles, and the risk value is calculated based on obstacle speed, direction, and size. Finally, a dynamic environment model with labeled risk levels is output, providing a perception basis for subsequent path planning; Distributed cooperative path planning layer: dynamically divide clusters based on robot position, task priority, and communication range, and manage intra-cluster cooperation by edge nodes. Based on environmental risk values, dynamically adjust the weights of path length, energy consumption, and safety, and generate a Pareto optimal path set through an improved multi-objective grey wolf optimizer. Combine A* algorithm to generate local candidate paths and filter out the optimal path that meets the multi-objective constraints for local real-time decision and global coordination fusion; Dynamic coordination conflict resolution layer: based on the current path of the robot and the trajectory of dynamic obstacles, predict the conflict probability in the next 1-2 steps through Monte Carlo tree search. When the conflict probability exceeds a threshold, trigger the distributed auction mechanism, and the robot generates a bid based on path cost, task priority, and remaining energy, with high-priority tasks receiving higher weights; Adjust the path through negotiation to avoid conflicts, and feed the results back to the path planning layer to form a closed-loop coordination. Continuous learning optimization layer: historical dynamic scene data is stored in the experience replay pool, and the meta-learning model is trained; when encountering a new scene, the model quickly adapts through a small number of gradient updates to generate an adaptive strategy; the optimized strategy and weight adjustment rule are injected into the pre-sequencing module to form an iterative optimization closed loop of "perception-planning-coordination-learning".

Images