Multi-agent obstacle avoidance collaborative decision-making method, device, equipment and medium

Abstract

The invention relates to a multi-agent obstacle avoidance collaborative decision-making method and device, equipment and a medium. The method comprises the following steps: acquiring individual attribute data, motion path data and motion state data of all agents and position, size and motion data of obstacles in real time through a pre-established communication network; predicting the future motion trail of the intelligent agent and the future state of the obstacle based on the data; calculating the minimum contact distance between any two intelligent agents and between each intelligent agent and the obstacle; when the minimum contact distance is smaller than a preset safety threshold value, marking the corresponding intelligent agent and all intelligent agents in the collaborative obstacle avoidance range as intelligent agents needing obstacle avoidance; generating an obstacle avoidance path through a path planning algorithm in combination with the future state of the obstacle; and calculating cooperative obstacle avoidance motion parameters based on the obstacle avoidance path, the agent individual attributes and the future motion trajectory, and generating a cooperative obstacle avoidance instruction. According to the method, autonomous cooperative obstacle avoidance of multiple agents in a dynamic environment is realized, and the security and cooperation efficiency are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent cooperative control, and in particular relates to multi-agent obstacle avoidance cooperative decision-making methods, devices, equipment and media. Background Technology

[0002] With the development of multi-agent cooperative control technology, obstacle avoidance cooperative decision-making technology for multi-agent systems in dynamic environments has emerged. This technology is characterized by its ability to handle the coordinated movement and collision avoidance of multiple agents (such as unmanned vehicles, robots, or drone swarms) in a shared space, and is widely used in logistics, intelligent manufacturing, and intelligent transportation. In traditional technologies, multi-agent obstacle avoidance typically employs two main strategies: one is a local path planning method based on independent decision-making, where each agent relies solely on its own sensor data (such as position and velocity) and simple rules (such as the artificial potential field method or A* algorithm) to calculate the obstacle avoidance path, lacking real-time information interaction with other agents; the other is a global optimization method based on centralized control, where a central server aggregates data from all agents and uses static algorithms (such as model predictive control or genetic algorithms) to generate a global obstacle avoidance scheme, but this requires processing high-dimensional data and issuing instructions. However, current methods have many problems: they cannot effectively predict the future state of obstacles (such as position and movement trend) and the interaction between agents in dynamic environments, increasing the risk of collisions; independent decision-making strategies ignore the scope of cooperative obstacle avoidance between agents, which can easily lead to group conflicts or path oscillations, while centralized control has poor real-time performance due to high computational latency and communication load, making it difficult to adapt to high-speed dynamic scenarios; existing methods do not consider the dynamic complexity of the environment (such as changes in the number of obstacles) and the urgency of system tasks (such as time constraints), and the safety threshold and cooperative scope are fixed, resulting in high obstacle avoidance risk, low cooperative motion efficiency, and insufficient overall system safety and cooperative performance. Summary of the Invention

[0003] Therefore, it is necessary to provide a multi-agent obstacle avoidance collaborative decision-making method, device, equipment, and medium that can solve the above problems.

[0004] Firstly, this application provides a multi-agent obstacle avoidance cooperative decision-making method, including:

[0005] The system acquires individual attribute data, motion path data, motion state data of each agent, as well as location data, size data, and motion data of obstacles through a pre-built communication network.

[0006] Based on the acquired data, a predictive model is used to predict the future motion trajectory of each agent and the future state of obstacles.

[0007] Based on future motion trajectories and future states, calculate the minimum contact distance between any two agents and between each agent and obstacles;

[0008] If any minimum contact distance is less than a preset safety threshold, the corresponding agent and all agents within the preset cooperative obstacle avoidance range will be marked as agents that need to avoid obstacles.

[0009] The system acquires the future state of obstacles, individual attribute data of the intelligent agent to be avoided, and future movement trajectory through a communication network. Based on the future state, individual attribute data, and future movement trajectory, a path planning algorithm is used to generate an obstacle avoidance path.

[0010] Based on the obstacle avoidance path, the individual attribute data of the intelligent agent to be avoided, and the future movement trajectory, the collaborative obstacle avoidance motion parameters are calculated, and the corresponding collaborative obstacle avoidance instructions are generated based on the collaborative obstacle avoidance motion parameters.

[0011] In one embodiment, a predictive model is used to predict the future motion trajectory of each agent and the future state of obstacles, including:

[0012] Based on the individual attribute data, motion path data and motion state data of the intelligent agents, the future motion trajectory of each intelligent agent is predicted through a time-series prediction model.

[0013] Based on the motion data of obstacles, obstacles are classified into dynamic obstacles or static obstacles;

[0014] For dynamic obstacles, based on their position and motion data, the future position and motion state are predicted by a state evolution model, and the future state of the obstacle is obtained by combining the future position and motion state with the size data.

[0015] For static obstacles, the position and size data are used as the future state of the obstacle;

[0016] Among them, the temporal prediction model includes long short-term memory network or temporal convolutional network, and the state evolution model includes Kalman filter or kinematic extrapolation model.

[0017] In one embodiment, calculating any two agents and the minimum contact distance between each agent and the obstacle includes:

[0018] Based on a pre-established kinematic-geometry model, the motion state of each agent and obstacle within a preset future time interval is dynamically modeled according to the future motion trajectory of the agent and the future state of the obstacle.

[0019] Based on the modeling results, the minimum Euclidean distance between any two agent size boundaries and between each agent size boundary and the obstacle size boundary is calculated in real time, and is taken as the minimum contact distance.

[0020] In one embodiment, an obstacle avoidance path is generated using a path planning algorithm, including:

[0021] Based on the future state of obstacles, the individual attribute data of the intelligent agents that need to avoid obstacles, and their future motion trajectories, an obstacle avoidance optimization function is constructed with obstacle avoidance risk and cooperative motion cost as optimization objectives. The obstacle avoidance risk is calculated based on the dynamic relationship between the future state of obstacles and the future motion trajectory of the intelligent agents that need to avoid obstacles, and the cooperative motion cost is obtained by weighting the Euclidean distance of the velocity vectors in the future motion trajectories of each intelligent agent.

[0022] Based on the improved distributed model predictive control algorithm, the obstacle avoidance optimization function is used to iteratively solve the obstacle avoidance path of each intelligent agent by satisfying the kinematic constraints and preset safety thresholds based on the individual attribute data of each agent.

[0023] In one embodiment, the cooperative obstacle avoidance motion parameters of the agent to be avoided are calculated, and corresponding cooperative obstacle avoidance instructions are generated based on the cooperative obstacle avoidance motion parameters, including:

[0024] Calculate the cooperative weight coefficient of the corresponding agent based on obstacle avoidance risk and cooperative movement cost;

[0025] Based on the obstacle avoidance path and the cooperative weight coefficient, the motion control parameters of the corresponding intelligent agent are calculated in real time.

[0026] Based on motion control parameters, collaborative obstacle avoidance commands are generated through a pre-set collaborative obstacle avoidance command library.

[0027] In one embodiment, the method further includes:

[0028] The number of obstacles is counted in real time during the dynamic modeling process;

[0029] The dynamic complexity of the environment is calculated based on the number and future state of obstacles and the future movement trajectories of each agent.

[0030] Acquire the preset task time limit data of each intelligent agent, and calculate the urgency of the system task by combining the corresponding motion path data;

[0031] Based on the dynamic complexity of the environment and the urgency of the system task, the preset safety threshold and / or preset collaborative obstacle avoidance range corresponding to the minimum contact distance are adjusted in real time according to preset adjustment rules.

[0032] In one embodiment, the individual attribute data includes at least one of size, shape, maximum speed, maximum acceleration, and minimum obstacle avoidance radius;

[0033] Motion path data includes at least one of the following: target location and planned waypoints;

[0034] The motion state data of each agent includes at least one of the following: current position, velocity, acceleration, and heading angle.

[0035] Secondly, this application also provides a multi-agent obstacle avoidance cooperative decision-making device, comprising:

[0036] The communication network module is used to acquire individual attribute data, motion path data, motion state data of each intelligent agent, as well as the position data, size data and motion data of obstacles through a pre-built communication network;

[0037] The trajectory prediction module is used to predict the future motion trajectory of each intelligent agent and the future state of obstacles based on the acquired data and a prediction model.

[0038] The safe distance calculation module is used to calculate the minimum contact distance between any two intelligent agents and between each intelligent agent and an obstacle, based on future movement trajectories and future states.

[0039] The collaborative obstacle avoidance decision module is used to mark the corresponding agent and all agents within the preset collaborative obstacle avoidance range as agents that need to avoid obstacles if any minimum contact distance is less than a preset safety threshold.

[0040] The path planning module is used to obtain the future state of obstacles, the individual attribute data of the intelligent agent to be avoided, and the future movement trajectory through the communication network, and to generate an obstacle avoidance path based on the future state, individual attribute data, and future movement trajectory using a path planning algorithm.

[0041] The collaborative control module is used to calculate collaborative obstacle avoidance motion parameters based on the obstacle avoidance path, the individual attribute data of the intelligent agent to be avoided, and the future motion trajectory, and to generate corresponding collaborative obstacle avoidance commands based on the collaborative obstacle avoidance motion parameters.

[0042] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described multi-agent obstacle avoidance cooperative decision-making method.

[0043] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described multi-agent obstacle avoidance cooperative decision-making method.

[0044] The aforementioned multi-agent obstacle avoidance collaborative decision-making method, device, computer equipment, and storage medium acquire individual attribute data, motion path data, and motion state data of each agent in real time through a pre-established communication network, as well as the position, size, and motion data of obstacles. Based on this data, it predicts the future motion trajectory of the agents and the future state of obstacles, thus mitigating the increased collision risk caused by the inability to effectively predict the future state of obstacles and the interaction between agents. It calculates the minimum contact distance between any two agents and between each agent and an obstacle. When the minimum contact distance is less than a preset safety threshold, the corresponding agent and all agents within the preset collaborative obstacle avoidance range are considered to be involved in the collision avoidance process. The agent is labeled as the agent requiring obstacle avoidance. By introducing a cooperative range mechanism, group conflicts or path oscillations caused by ignoring the cooperative obstacle avoidance range under independent decision-making strategies are avoided. At the same time, real-time data processing through the communication network reduces the high computational latency and communication load of centralized control, improving the system's real-time performance. Obstacle avoidance paths are generated through path planning algorithms, and combined with the individual attribute data and future movement trajectory of the agents requiring obstacle avoidance, cooperative obstacle avoidance motion parameters are calculated to generate instructions, optimizing obstacle avoidance risk and cooperative motion efficiency. Autonomous cooperative obstacle avoidance is achieved in dynamic environments, improving safety and collaborative performance, and solving the problems of high obstacle avoidance risk and low collaborative efficiency caused by fixed thresholds and ranges. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the multi-agent obstacle avoidance collaborative decision-making method of the present invention;

[0047] Figure 2 This is a structural diagram of the multi-agent obstacle avoidance collaborative decision-making device of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] In one embodiment, such as Figure 1As shown, a multi-agent obstacle avoidance collaborative decision-making method is provided. This embodiment illustrates the application of this method to a multi-agent cluster system (such as unmanned vehicle platoons, drone swarms, or industrial robot groups). This method is also applicable to server-centralized control architectures (such as cloud scheduling platforms) or terminal-server collaborative systems (where the terminal is the agent's local control unit and the server is an edge computing node). The implementation environment hardware architecture comprises a three-layer structure: the perception layer uses sensors such as positioning modules (GPS / BeiDou), inertial measurement units, and lidar mounted on the agents to collect motion state data (position, velocity, acceleration) and raw obstacle data in real time; the communication layer can rely on a pre-built 5G / V2X low-latency network to achieve sharing of individual attributes, planned paths, and real-time status data among agents, while broadcasting obstacle position / size / motion data to the decision node; the decision layer, based on the terminal embedded processor or GPU server cluster, runs trajectory prediction, safe distance calculation, and collaborative path planning algorithms.

[0050] Taking intelligent logistics warehousing as a typical application scenario: When an AGV fleet is transporting goods in a dynamic shelf environment, the system obtains the path planning data (including target position and waypoints) and motion status (heading angle, acceleration) of all AGVs through the communication layer, and integrates the data of moving forklifts (dynamic obstacles) detected by the vision system. If the prediction model determines that the minimum contact distance between an AGV and a forklift will fall below the 0.5-meter safety threshold in 2 seconds, then that AGV and other AGVs within a 3-meter radius of the collaborative obstacle avoidance range are marked as intelligent agents requiring obstacle avoidance. The path planning module combines the maximum acceleration constraints of each AGV and the future state of the forklift to generate a collaborative obstacle avoidance path, and generates a synchronous steering instruction set through motion control parameters. In intelligent manufacturing scenarios, industrial robotic arm groups interact through device architecture: when robotic arm A triggers an alarm for the minimum contact distance with the high-speed conveyor belt (dynamic obstacle), the collaborative obstacle avoidance decision module links adjacent robotic arms to adjust their motion paths, and coordinates the acceleration parameters of multiple arms through a distributed model predictive control algorithm to maintain production line synchronization while ensuring a safe distance.

[0051] In this embodiment, the method includes the following steps:

[0052] S01 acquires individual attribute data, motion path data, motion state data of each intelligent agent, as well as the location data, size data, and motion data of obstacles through a pre-built communication network.

[0053] Individual attribute data comprises the inherent physical and motion constraints of the agent, used to characterize its physical boundaries and dynamic limits; motion path data consists of planned motion target information, used to guide the agent's global navigation; motion state data is real-time dynamic information, used to reflect the agent's instantaneous motion behavior; obstacle position data, size data, and motion data correspond to its spatial coordinates, geometric dimensions (e.g., length, width, and height), and motion trends (e.g., velocity vector), respectively. In implementation, a pre-built communication network architecture can be used (adopting a three-tier architecture of core network-edge computing nodes-agent terminal: the core network is responsible for global data aggregation (e.g., task time limits, global paths), edge nodes (deployment spacing ≤ 1km) are responsible for low-latency data forwarding within the area (communication latency ≤ 100ms), and the agent terminal is equipped with a communication module (supporting 5G NR-V2X protocol); the network adopts DDS (Data Distribution System). The communication protocol uses a Service (QoS) configuration of 'reliable transmission + low latency priority'. A single edge node supports concurrent communication of ≥50 agents, with a single agent transmitting ≤1kb of data per transmission, including compressed data of individual attributes and motion states. The perception layer collects raw data in real time through sensors (such as positioning modules, inertial measurement units, and lidar) on the agents. The communication layer shares and broadcasts this data (including individual attributes, path planning, and real-time status) across multiple nodes. Obstacle data is synchronously uploaded to the network by an environmental monitoring system (such as a vision or radar system). The data acquisition process utilizes distributed communication protocols (such as MQTT or DDS) for efficient transmission, providing a real-time input foundation for subsequent trajectory prediction and collaborative obstacle avoidance.

[0054] S02, based on the acquired data, a prediction model is used to predict the future movement trajectory of each agent and the future state of obstacles.

[0055] The future trajectory is the expected position sequence of the agent within a preset time interval, derived from the agent's real-time motion state and path data. The future state is the predicted geometric position, size, and dynamic behavior of obstacles. In practice, based on the agent's individual attribute data (such as size and maximum acceleration), path data (such as target position), and motion state data (such as heading angle and current speed), a temporal prediction model (including long short-term memory networks or temporal convolutional networks) is used to analyze historical motion sequences and generate the future trajectory of each agent. Based on the obstacle's motion data (such as velocity vector), it is classified as a dynamic obstacle or a static obstacle. For dynamic obstacles, based on their position and motion data, a state evolution model (such as Kalman filtering or kinematic extrapolation model) can be used to predict the future position and motion state, and the size data is fused to output the complete future state. For static obstacles, the position and size data are directly used as the future state.

[0056] S03, based on future motion trajectories and future states, calculate the minimum contact distance between any two agents and between each agent and obstacles.

[0057] The minimum contact distance is the minimum Euclidean distance between the boundaries of the agent's dimensions and the boundaries of the obstacle's dimensions, which can be calculated in real time based on a pre-established kinematic-geometric model. In practice, the system dynamically models the motion state of each agent and obstacle within a preset future time interval based on the agent's future trajectory and the obstacle's future state. It generates a time-series spatial occupancy model using kinematic equations combined with geometric boundary data. Based on the modeling results, it calculates in real time the minimum Euclidean distance between any two agent boundaries and the minimum Euclidean distance between each agent boundary and the obstacle boundary, outputting this distance as the minimum contact distance. This addresses the collision risk caused by neglecting kinematic-geometric interactions in traditional methods.

[0058] S04. If any minimum contact distance is less than a preset safety threshold, then the corresponding agent and all agents within the preset cooperative obstacle avoidance range are marked as agents that need to avoid obstacles.

[0059] If any minimum contact distance is less than a preset safety threshold (this threshold is the critical value for collision risk, and the initial value is set by the individual attributes of the agent; for example, for a small drone (body radius ≤ 0.3m), the threshold is the drone's body radius + the equivalent radius determined by the obstacle size + 0.1m (aerodynamic disturbance margin), and can be adjusted in real time according to the dynamic complexity of the environment and the urgency of the system task), the corresponding agent (i.e., the agent whose distance exceeds the preset safety threshold) and its surroundings within the preset cooperative obstacle avoidance range (radius R, preset rule: R = agent's maximum braking distance × 1.5 + 0.5m; where, agent's maximum braking distance = (agent's current speed) / (2 × agent's maximum deceleration); if the current speed is 0, then take...) All agents whose maximum speed corresponds to the braking distance (and R can be dynamically optimized in subsequent steps) are marked as agents requiring obstacle avoidance (the set of agents requiring cooperative path adjustment, including agents directly facing risks and their neighboring collaborators). In implementation, the minimum contact distance data is compared in real-time with a preset safety threshold. If the distance is less than the preset safety threshold, based on the agent's current location data (e.g., GPS coordinates) and a preset cooperative obstacle avoidance range, all agents within the range can be quickly retrieved using a spatial query algorithm (e.g., KD-Tree or range search). The target agent and its retrieved agent ID are marked as agents requiring obstacle avoidance, and this marking data is broadcast through a communication network to provide input for subsequent obstacle avoidance path generation.

[0060] S05: Obtain the future state of the obstacle, the individual attribute data of the intelligent agent to be avoided, and the future movement trajectory through the communication network, and generate an obstacle avoidance path based on the future state, individual attribute data, and future movement trajectory using a path planning algorithm.

[0061] In implementation, an obstacle avoidance optimization function can be constructed based on the future state of the obstacle, the individual attribute data of the agent requiring obstacle avoidance, and its future motion trajectory. The function aims to optimize obstacle avoidance risk and cooperative motion cost. The obstacle avoidance risk is dynamically calculated based on the future state and future motion trajectory of the obstacle, while the cooperative motion cost is obtained by weighted calculation of the Euclidean distance of the velocity vectors in the future motion trajectories of each agent. Based on an improved distributed model predictive control algorithm, the obstacle avoidance optimization function is used to iteratively solve the problem under the condition that the kinematic constraints (such as maximum acceleration limit) of the individual attribute data of the agent are met and the preset safety threshold is used as the stopping condition, thereby generating the obstacle avoidance path for each agent requiring obstacle avoidance.

[0062] S06. Based on the obstacle avoidance path, the individual attribute data of the intelligent agent to be avoided, and the future movement trajectory, calculate the cooperative obstacle avoidance motion parameters, and generate the corresponding cooperative obstacle avoidance command based on the cooperative obstacle avoidance motion parameters.

[0063] In implementation, the system can calculate the collaborative weight coefficient (reflecting the decision weight of the agent in collaborative obstacle avoidance, such as high weight corresponding to high priority path adjustment) in real time based on the obstacle avoidance path, individual attribute data of the agent to be avoided, and future movement trajectory, as well as the obstacle avoidance risk and collaborative movement cost. Based on the obstacle avoidance path (such as time-position sequence) and collaborative weight coefficient, the system solves the motion control parameters (such as target speed, target acceleration, and heading angle, constrained by the maximum speed and minimum obstacle avoidance radius in the individual attribute data) in real time through multi-objective constraints. Through a pre-set collaborative obstacle avoidance command library (storing standard command templates, such as the mapping relationship of "turn left 30 degrees, decelerate to 0.5m / s"), the motion control parameters are converted into collaborative obstacle avoidance commands, which are then sent to the corresponding agents through a pre-built communication network. This optimizes the command generation efficiency, solves the path oscillation problem caused by command delay in traditional methods, and improves the overall security and collaborative performance of the system.

[0064] The aforementioned multi-agent obstacle avoidance collaborative decision-making method acquires individual attribute data, motion path data, motion state data of each agent, and the position, size, and motion data of obstacles in real time through a pre-built communication network, ensuring the real-time and comprehensiveness of the data source. Based on this data, a predictive model is used to predict the future motion trajectory of the agents and the future state of obstacles, solving the problem of increased collision risk caused by the difficulty in predicting the future state of obstacles and the interaction influence of agents in traditional methods. Based on the prediction results, the minimum contact distance between any two agents and between each agent and obstacles is calculated. When the minimum contact distance is less than a preset safety threshold, the corresponding agent and all agents within the preset collaborative obstacle avoidance range are marked as agents that need to avoid obstacles. An obstacle avoidance path is generated by combining the path planning algorithm with the future state of obstacles. For example, an improved distributed model predictive control algorithm is used to construct an optimization function. Through obstacle avoidance risk and collaborative motion cost constraints, the obstacle avoidance path is ensured to be efficient and conform to agent kinematic constraints (such as maximum acceleration limits), reducing the high computational latency of centralized control. Based on the obstacle avoidance path and agent dynamic data, collaborative obstacle avoidance motion parameters are calculated and instructions are generated. Synchronous control commands are quickly issued through a preset instruction library to optimize collaborative efficiency. To achieve autonomous and collaborative obstacle avoidance in dynamic environments, thereby improving system security and collaborative performance.

[0065] In one embodiment, a predictive model is used to predict the future motion trajectory of each agent and the future state of obstacles, including:

[0066] S11, based on the individual attribute data, motion path data and motion state data of the intelligent agent, predicts the future motion trajectory of each intelligent agent through a time-series prediction model;

[0067] S12, based on the motion data of the obstacles, classify the obstacles into dynamic obstacles or static obstacles;

[0068] S13. For dynamic obstacles, based on their position data and motion data, the future position and motion state are predicted through a state evolution model, and the future state of the obstacle is obtained by combining the future position and motion state with the size data.

[0069] S14, For static obstacles, the position data and size data are used as the future state of the obstacle;

[0070] Among them, the temporal prediction model includes long short-term memory network or temporal convolutional network, and the state evolution model includes Kalman filter or kinematic extrapolation model.

[0071] Specifically, based on the agent's individual attribute data, motion path data, and motion state data, a temporal prediction model (Long Short-Term Memory Network LSTM or Temporal Convolutional Network TCN) can be used to process the agent's historical motion sequence. If it is a Long Short-Term Memory (LSTM) network, a 3-layer structure is adopted: input layer neurons = 6 (corresponding to the agent's 6-dimensional features of 'current position x / y, velocity x / y, acceleration x / y'), hidden layer neurons = 32 (activation function is ReLU), output layer neurons = 2×T (T is the prediction time step, default T = 5, each step interval is 0.5s, outputting the x / y coordinates of the next 5 time steps); the model training data is the agent's historical motion data (collection frequency 10Hz, sample size ≥ 10000), the loss function is the mean squared error (MSE) between the predicted coordinates and the actual coordinates, and the training optimizer is Adam (learning rate 0.001); if it is a Temporal Convolutional Network (TCN), two 1D convolutional layers (convolution kernel size 3, stride 1) and one pooling layer (pooling kernel size 2) are adopted, the output format is the same as LSTM, and the training parameters are the same as LSTM. In implementation, the temporal prediction model takes as input motion state data of the agent (e.g., velocity and heading angle of 10 consecutive frames) superimposed with motion path data (e.g., remaining path points). It learns motion patterns through a neural network and outputs a future position sequence arranged by timestamps at a fixed future time (e.g., 5 seconds, determined based on configurable computing resources), forming the future motion trajectory. A binary judgment is performed based on the vector magnitude (i.e., velocity) in the obstacle motion data: if the velocity consistently exceeds a threshold (e.g., 0.2 m / s), it is marked as a dynamic obstacle; otherwise, it is marked as a static obstacle. For dynamic obstacles, a state evolution model is invoked: such as a Kalman filter model, suitable for scenarios with motion noise (e.g., scenarios with pedestrians in the path), through a state equation (position-velocity relationship) x k =A·x k-1 +B·u k +w k , where x k =[p x p y v x v y ] T , ( Let k be the position at time k. Let A be the velocity at time k), and let A be the state transition matrix (4×4, with diagonal elements of 1, A[0][2]=A[1][3]=Δ). t Δ t =0.1s0 is the sampling interval), B is the control matrix (4×2, B[2][0]=B[3][1]=Δ t ), u k =[a x a y ] T(Acceleration at time k), w k Let z be the process noise (Gaussian distribution, variance diag([0.01,0.01,0.001,0.001])); and z be the observation equation (sensor data). k =H·x k +v k Where z k =[p xobs p yobs ] T Sensor observation position), H is the observation matrix (2×4, H[0][0]=H[1][1]=1, the rest are 0), v k To observe the noise (Gaussian distribution, variance diag([0.02,0.02]), iteratively predict future position and velocity; if it is a kinematic extrapolation model: use the uniform acceleration motion formula p t +Δ t = Where p t Let v be the position at time t. t Let a be the velocity at time t. t Let Δ be the acceleration at time a (if the acceleration is unknown, it is assumed to be 0), and Δ be the acceleration at time a. t =0.1s, extrapolation time range ≤3s, and generate a future state envelope with geometric boundaries by combining the size data. For static obstacles, the current position data and size data are directly output as the future state, without the need for prediction calculations.

[0072] In one embodiment, calculating any two agents and the minimum contact distance between each agent and the obstacle includes:

[0073] S21, based on the pre-established kinematic-geometric model, dynamically model the motion state of each agent and obstacle within a preset future time interval according to the future motion trajectory of the agent and the future state of the obstacle;

[0074] S22, based on the modeling results, calculate in real time the minimum Euclidean distance between any two agent size boundaries and between each agent size boundary and the obstacle size boundary, as the minimum contact distance.

[0075] For example, based on a pre-established kinematic-geometry model (coupling the kinematic and geometric models), Newton's second law is used to derive the relationship between the position of the agent / obstacle and time, such as Newton's equations of motion. With geometric models: Agents / obstacles are abstracted as convex polygons (2D scene) or convex polyhedra (3D scene). For example, a drone is abstracted as a rectangle (length × width = 1.2m × 0.8m), and a circular obstacle is abstracted as a circle (radius r = 0.5m). The coupling method is to use the p output of the kinematic sub-model. tUsing the central coordinates of the geometric model, the spatial occupancy area of ​​the agent / obstacle at time t is obtained; based on the future motion trajectory of the agent and the future state of the obstacle, the space occupied by each agent and obstacle in the preset future time interval (the setting rule is T) is determined. p =max(maximum braking time of agent, maximum movement time of obstacle), where, maximum braking time of agent = current speed of agent / maximum deceleration of agent, maximum movement time of obstacle = 3s (if the obstacle is dynamic and the movement state is unstable, take a fixed upper limit of 3s); the default value range is [1s, 5s], with a step size of 0.1s. ) to dynamically model the motion state within the range. This modeling process constructs a time-continuous spatial occupancy volume by solving the motion differential equation and combining it with a geometric bounding box (such as an OBB oriented bounding box); based on this modeling result, the minimum Euclidean distance between any two agent size boundaries is calculated in real time (e.g., by solving the closest point between convex bodies using the Gilbert-Johnson-Keerthi algorithm, such as: Step 1. Construct a convex hull, abstracting the dynamic geometric model of the agent / obstacle (at time t) into a convex hull, such as the vertices of the rectangular convex hull of the UAV being (p x ±0.6,p y ±0.4), the convex hull of the circular obstacle is a circle (approximated by an inscribed regular hexagon with 6 vertices); Step 2. Initialize the simplex set (the set of points for iteration), and set the search direction (default from the center of the agent's convex hull to the center of the obstacle's convex hull); Step 3. Iteratively solve, in each iteration, select the point closest to the origin from the convex hull and add it to the simplex, update the search direction, until the simplex contains the origin (indicating a collision, the distance is 0) or the search direction cannot find a closer point (at this time, the distance between the two points in the simplex is the minimum Euclidean distance); Time series calculation, for t∈[0,T p Repeat steps 1-3 at each time step, taking the minimum distance across all time steps as the minimum contact distance, and the minimum Euclidean distance between the boundaries of each agent's size and the obstacle's size (sweptvolume detection can be used for dynamic obstacles), and outputting this distance as the minimum contact distance. This precise quantification of collision risk at the geometric kinematic level solves the collision problems caused by neglecting kinematic-geometric interactions in traditional methods.

[0076] In one embodiment, an obstacle avoidance path is generated using a path planning algorithm, including:

[0077] S31. Based on the future state of the obstacle, the individual attribute data of the intelligent agent to be avoided, and the future motion trajectory, an obstacle avoidance optimization function is constructed with obstacle avoidance risk and cooperative motion cost as optimization objectives. The obstacle avoidance risk is calculated based on the dynamic relationship between the future state of the obstacle and the future motion trajectory of the intelligent agent to be avoided, and the cooperative motion cost is obtained by weighting the Euclidean distance of the velocity vectors in the future motion trajectories of each intelligent agent.

[0078] S32, based on an improved distributed model predictive control algorithm, utilizes an obstacle avoidance optimization function to iteratively solve the obstacle avoidance path for each intelligent agent that needs to avoid obstacles, by satisfying the kinematic constraints and preset safety thresholds defined by the individual attribute data of each agent.

[0079] Specifically, an obstacle avoidance optimization function can be constructed based on the future state of obstacles, the individual attribute data of the agents to be avoided, and their future motion trajectories. The function aims to optimize obstacle avoidance risk and cooperative motion cost. Obstacle avoidance risk is a potential collision probability index between the agent and the obstacle, calculated through the dynamic relationship between the future state of the obstacle and the future motion trajectory of the agent to be avoided (e.g., the rate of change of relative position and the reciprocal of the minimum Euclidean distance), quantifying the safety threat in dynamic interactions. Cooperative motion cost characterizes the loss of efficiency in multi-agent cooperative motion, obtained by weighted summation of the Euclidean distances of the velocity vectors in the future motion trajectories of each agent (e.g., weighted calculation based on the magnitude of velocity differences, reflecting cooperative consistency). The obstacle avoidance optimization function formula is min... j =w R ×R total +w c ×C total , where R total Total obstacle avoidance risk R total =Σ(R i ), R i The obstacle avoidance risk of the i-th intelligent agent that needs to avoid obstacles. d i Let C be the minimum contact distance between the i-th agent and the nearest obstacle, ε = 0.01m to avoid the denominator being zero. total For the total cost of coordinated motion, C total =ΣΣ(|‖v i (t)-v j (t)‖2×w ij )), ‖·‖2 is the Euclidean distance, v i (t) is the velocity vector of the i-th agent at time t, w ij The cooperative weights for agents i and j (if i and j are within the cooperative obstacle avoidance range, w) ij =1, otherwise w ij =0)w R and w c For the target weight, satisfying wR +w c =1, default w R =0.7 (prioritizing safety), w c =0.3 (suboptimal guarantee of collaborative efficiency), which can be adjusted according to the urgency of the task (e.g., w when the task is urgent). c Increased to 0.5).

[0080] Based on an improved distributed model predictive control algorithm (a distributed optimization framework that achieves global collaboration through local decision-making and information exchange), the obstacle avoidance optimization function is used to solve the obstacle avoidance path of each agent in real time, under the condition of satisfying the kinematic constraints defined by individual attribute data (such as the agent's maximum speed, maximum acceleration, maximum heading angle, and maximum obstacle avoidance radius constraints) and a preset safety threshold as the stopping condition for iteration. The improved distributed model predictive control algorithm uses a 'neighbor interaction' communication topology, where each agent only exchanges predicted trajectories with agents within the collaborative obstacle avoidance range (requiring no global communication and reducing load); it optimizes the time domain, predicting the time domain N... p =T p Δt (Δt = 0.1s, such as T) p =2s N p =20), control time domain N c =N p 2 (i.e., only optimize the control variables in the first 10 steps, and use the control variables from step 10 in the last 10 steps), iterative steps: Step 1. Each agent that needs to avoid obstacles initializes the predicted trajectory based on local data (its own attributes, the future state of the obstacle); Step 2. The agent exchanges predicted trajectories with its neighbors and calculates J (the obstacle avoidance optimization function value); Step 3. If J > a preset threshold (e.g., 10), the control variables (acceleration, heading angle) are updated using gradient descent, and the predicted trajectory is adjusted; Steps 2-3 are repeated until J ≤ 10 or the number of iterations ≥ 10, and the final predicted trajectory is output as the obstacle avoidance path; By repeatedly optimizing the path point sequence within a preset time step, the obstacle avoidance path minimizes the collaborative cost while reducing risk, thereby improving the efficiency and safety of autonomous decision-making in dynamic environments.

[0081] In one embodiment, the cooperative obstacle avoidance motion parameters of the agent to be avoided are calculated, and corresponding cooperative obstacle avoidance instructions are generated based on the cooperative obstacle avoidance motion parameters, including:

[0082] S41, Calculate the cooperative weight coefficient of the corresponding agent based on obstacle avoidance risk and cooperative motion cost;

[0083] S42, based on the obstacle avoidance path and the cooperative weight coefficient, calculates the motion control parameters of the corresponding intelligent agent in real time;

[0084] S43 generates collaborative obstacle avoidance commands based on motion control parameters and a preset collaborative obstacle avoidance command library.

[0085] For example, based on obstacle avoidance risk (i.e., the probability index of potential collision between the agent and the obstacle) and cooperative motion cost (a loss measure of the efficiency of multi-agent cooperative motion), the corresponding agent's cooperative weight coefficient is calculated (characterizing the priority parameter of the agent in cooperative obstacle avoidance decision-making, the value of which is proportional to the obstacle avoidance risk and inversely proportional to the cooperative motion cost). w i Let k be the collaborative weight coefficient of the i-th agent. R k c This is the scaling factor, with k as the default value. R =10, k c =1, which can be adjusted according to the type of intelligent agent, such as drone k R =15); Based on the obstacle avoidance path and cooperative weight coefficients, the motion control parameters of the corresponding agent are calculated in real time (including target velocity, target acceleration, and heading angle adjustment). The calculation process needs to meet the constraints of maximum velocity, maximum acceleration, and minimum obstacle avoidance radius in the individual attribute data. Model predictive control (MPC) can be used to calculate the motion control parameters (target velocity, target acceleration, target heading angle). The steps are as follows: Input the reference trajectory of the obstacle avoidance path and the cooperative weight coefficients w i The process involves: constructing the MPC objective function; substituting the kinematic constraints to solve for the target acceleration of the optimal control quantity, integrating to obtain the target velocity, and determining the target heading angle from the tangent direction of the obstacle avoidance path; outputting a set of (target velocity, target acceleration, target heading angle) every 0.1s; converting motion control parameters into executable cooperative obstacle avoidance commands (such as discrete command sets like "decelerate to 0.8m / s" or "turn left 15 degrees") through a pre-built cooperative obstacle avoidance command library (a database storing standardized command mapping relationships, such as a velocity-command mapping table or a heading angle-turning command mapping table), and distributing these commands to the agent requiring obstacle avoidance via a pre-built communication network. By dynamically adjusting control parameters through quantified cooperative weights and combining this with the command library's fast mapping mechanism, the system optimizes command generation efficiency while ensuring obstacle avoidance safety, thereby enhancing the dynamic cooperative capability of the multi-agent system. In intelligent logistics scenarios, when an AGV cluster needs to coordinate to avoid a moving forklift, the coordination weight coefficient of each AGV is calculated based on the forklift approach risk (obstacle avoidance risk) and the speed difference of the AGV queue (coordinated motion cost). Based on this coefficient, a larger acceleration adjustment is generated for high-priority AGVs and mapped to an "emergency deceleration 30%" instruction through the instruction library, while a small heading angle correction instruction is generated for low-priority AGVs to achieve differentiated coordinated obstacle avoidance control.

[0086] In one embodiment, the method further includes:

[0087] S51 counts the number of obstacles in real time during dynamic modeling;

[0088] S52 calculates the dynamic complexity of the environment based on the number and future states of obstacles and the future motion trajectories of each agent.

[0089] S53, acquire the preset task time limit data of each intelligent agent, and calculate the urgency of the system task in combination with the corresponding motion path data;

[0090] S54, based on the dynamic complexity of the environment and the urgency of the system task, adjusts the preset safety threshold and / or preset cooperative obstacle avoidance range corresponding to the minimum contact distance in real time according to preset adjustment rules.

[0091] Specifically, during the dynamic modeling process (simulating continuous spatial occupancy of the future motion states of agents and obstacles based on a pre-established kinematic-geometric model), the number of obstacles is counted in real time (i.e., the total number of obstacles detected in the environment, which can be updated in real time through sensor data); based on the number of obstacles and the future motion trajectories of each agent, the dynamic complexity of the environment is calculated (the quantitative index D of environmental instability, obtained by multiplying the obstacle density factor and the agent motion perturbation coefficient, such as the number of obstacles D). obs Multiply by the rate of change D of the agent's future trajectory over a period of time. agent :D=D obs ×D agent Simultaneously, acquire the preset task time limit data for each agent (i.e., the deadline constraint for the agent to complete the task), and combine it with the corresponding motion path data (such as the remaining path length) to calculate the system task urgency (a measure of the urgency of task completion, calculated by weighting the ratio of the reciprocal of the time limit to the path completion rate: u = (1 / T) re )×(L co / L to )×k u1 ), T re L represents the remaining time for the agent's task (obtained by subtracting the current time from the deadline when the task instruction was issued, in seconds); co L represents the length of the path already completed by the agent (the distance from the current position to the starting point of the originally planned path, in meters); to k represents the total length of the originally planned path (the distance from the starting point to the target location, in meters). u This is the task type coefficient; based on the dynamic complexity of the environment and the urgency of the system task, according to preset adjustment rules (such as lookup table method), the preset safety threshold (initial collision risk threshold, such as increasing from 0.5 meters to 0.8 meters to cope with high complexity) and / or preset cooperative obstacle avoidance range (initial cooperative radius, such as expanding from 3 meters to 5 meters to adapt to high urgency) corresponding to the minimum contact distance are adjusted in real time, dynamically optimizing obstacle avoidance decisions to adapt to environmental changes and improve robustness and efficiency in variable scenarios.

[0092] In one embodiment, S61, the individual attribute data includes at least one of size, shape, maximum speed, maximum acceleration, and minimum obstacle avoidance radius;

[0093] Motion path data includes at least one of the following: target location and planned waypoints;

[0094] The motion state data of each agent includes at least one of the following: current position, velocity, acceleration, and heading angle.

[0095] For example, individual attribute data consists of the inherent physical and motion constraints of the agent, including at least one of the following: size (e.g., geometric dimensions of length, width, and height), shape (e.g., rectangular or circular outline), maximum speed (upper limit of motion speed), maximum acceleration (limit of acceleration change), and minimum obstacle avoidance radius (minimum turning radius required for safe obstacle avoidance); motion path data consists of planned motion target information, including at least one of the following: target position (coordinates of the endpoint that the agent needs to reach) and planned path points (a sequence of key positions in a preset path); the motion state data of each agent is real-time dynamic information, including current position (e.g., GPS coordinates), speed (instantaneous) and velocity. The data includes at least one of the following: velocity vector, acceleration (rate of change of velocity), and heading angle (angle of motion direction). This data is generated through a pre-built communication network module. The communication layer relies on a network architecture (such as 5G / V2X) and uses sensors (such as positioning modules, inertial measurement units, and lidar) on the intelligent agent to collect raw data in real time. The data is shared and broadcast to multiple nodes through distributed communication protocols (such as MQTT or DDS). The environmental monitoring system (such as vision or radar system) simultaneously uploads obstacle-related information to ensure the real-time and comprehensiveness of the data source, providing a structured input basis for subsequent trajectory prediction and collaborative obstacle avoidance decision-making.

[0096] The aforementioned multi-agent obstacle avoidance collaborative decision-making method acquires individual attribute data, motion path data, and motion state data of each agent in real time through a pre-built communication network, as well as the position, size, and motion data of obstacles. This addresses the collision risk problem caused by data isolation and poor real-time performance in traditional methods. Based on this data, a predictive model (such as a Long Short-Term Memory network or Kalman filter) is used to predict the future motion trajectory of the agents and the future state of obstacles (including the positional evolution of dynamic obstacles and the fixed state of static obstacles). This overcomes the difficulty in predicting the future state of obstacles and the interaction between agents, reducing potential collision risks. The method calculates the minimum contact distance between any two agents and between each agent and an obstacle using a kinematic-geometric model. When this distance is less than a preset safety threshold, the corresponding agent and all agents within the preset collaborative obstacle avoidance range are marked as agents requiring obstacle avoidance, introducing a collaborative range mechanism to resolve the issue. Independent decision-making strategies ignore group conflicts or path oscillations caused by the scope of collaborative obstacle avoidance. A path planning algorithm is used to construct a function with obstacle avoidance risk and collaborative movement cost as optimization objectives, combining the future state of obstacles, individual attribute data of the agents to be avoided, and future movement trajectories. This generates obstacle avoidance paths. A distributed architecture reduces the computational latency and communication load of centralized control, improving real-time performance. Based on the obstacle avoidance paths, individual attribute data, and future movement trajectories, collaborative obstacle avoidance motion parameters (collaborative weight coefficients and motion control parameters) are calculated to generate collaborative obstacle avoidance commands. Safety thresholds and collaborative obstacle avoidance ranges are dynamically adjusted according to the dynamic complexity of the environment (calculated based on the number of obstacles and agent movement trajectories) and the urgency of the system task (combined with task time limit data and movement path data), optimizing obstacle avoidance risk and collaborative efficiency. This enables autonomous collaborative obstacle avoidance by multiple agents in dynamic environments, improving safety, collaborative performance, and system robustness.

[0097] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0098] Based on the same inventive concept, this application also provides a multi-agent obstacle avoidance cooperative decision-making device for implementing the multi-agent obstacle avoidance cooperative decision-making method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the multi-agent obstacle avoidance cooperative decision-making device provided below can be found in the limitations of the multi-agent obstacle avoidance cooperative decision-making method described above, and will not be repeated here.

[0099] In one exemplary embodiment, such as Figure 2 As shown, a multi-agent obstacle avoidance collaborative decision-making device is provided, comprising:

[0100] The communication network module 101 is used to acquire individual attribute data, motion path data, motion state data of each intelligent agent, as well as the position data, size data and motion data of obstacles through a pre-built communication network;

[0101] The trajectory prediction module 102 is used to predict the future motion trajectory of each intelligent agent and the future state of obstacles based on the acquired data and a prediction model.

[0102] The safe distance calculation module 103 is used to calculate the minimum contact distance between any two intelligent agents and between each intelligent agent and an obstacle based on the future motion trajectory and future state.

[0103] The collaborative obstacle avoidance decision module 104 is used to mark the corresponding intelligent agent and all intelligent agents within the preset collaborative obstacle avoidance range as intelligent agents that need to avoid obstacles if any minimum contact distance is less than a preset safety threshold.

[0104] The path planning module 105 is used to obtain individual attribute data, motion path data and motion state data of the intelligent agent to be avoided through the communication network, and generate an obstacle avoidance path by combining the future state of the obstacle with the path planning algorithm.

[0105] The collaborative control module 106 is used to calculate collaborative obstacle avoidance motion parameters based on the obstacle avoidance path, the individual attribute data of the intelligent agent to be avoided, and the future motion trajectory, and to generate corresponding collaborative obstacle avoidance commands based on the collaborative obstacle avoidance motion parameters.

[0106] In one embodiment, the trajectory prediction module 102 is further configured to:

[0107] Based on the individual attribute data, motion path data and motion state data of the intelligent agents, the future motion trajectory of each intelligent agent is predicted through a time-series prediction model.

[0108] Based on the motion data of obstacles, obstacles are classified into dynamic obstacles or static obstacles;

[0109] For dynamic obstacles, based on their position and motion data, the future position and motion state are predicted by a state evolution model, and the future state of the obstacle is obtained by combining the future position and motion state with the size data.

[0110] For static obstacles, the position and size data are used as the future state of the obstacle;

[0111] Among them, the temporal prediction model includes long short-term memory network or temporal convolutional network, and the state evolution model includes Kalman filter or kinematic extrapolation model.

[0112] In one embodiment, the safe distance calculation module 103 is further configured to:

[0113] Based on a pre-established kinematic-geometry model, the motion state of each agent and obstacle within a preset future time interval is dynamically modeled according to the future motion trajectory of the agent and the future state of the obstacle.

[0114] Based on the modeling results, the minimum Euclidean distance between any two agent size boundaries and between each agent size boundary and the obstacle size boundary is calculated in real time, and is taken as the minimum contact distance.

[0115] In one embodiment, the path planning module 105 is further configured to:

[0116] Based on the future state of obstacles, the individual attribute data of the intelligent agents that need to avoid obstacles, and their future motion trajectories, an obstacle avoidance optimization function is constructed with obstacle avoidance risk and cooperative motion cost as optimization objectives. The obstacle avoidance risk is calculated based on the dynamic relationship between the future state of obstacles and the future motion trajectory of the intelligent agents that need to avoid obstacles, and the cooperative motion cost is obtained by weighting the Euclidean distance of the velocity vectors in the future motion trajectories of each intelligent agent.

[0117] Based on the improved distributed model predictive control algorithm, the obstacle avoidance optimization function is used to iteratively solve the obstacle avoidance path of each intelligent agent by satisfying the kinematic constraints and preset safety thresholds based on the individual attribute data of each agent.

[0118] In one embodiment, the cooperative control module 106 is further configured to:

[0119] Calculate the cooperative weight coefficient of the corresponding agent based on obstacle avoidance risk and cooperative movement cost;

[0120] Based on the obstacle avoidance path and the cooperative weight coefficient, the motion control parameters of the corresponding intelligent agent are calculated in real time.

[0121] Based on motion control parameters, collaborative obstacle avoidance commands are generated through a pre-set collaborative obstacle avoidance command library.

[0122] In one embodiment, the trajectory prediction module 102 is further configured to:

[0123] The number of obstacles is counted in real time during the dynamic modeling process;

[0124] The number and future state of obstacles, the future motion trajectory of each intelligent agent, and the dynamic complexity of the environment are calculated.

[0125] Acquire the preset task time limit data of each intelligent agent, and calculate the urgency of the system task by combining the corresponding motion path data;

[0126] Based on the dynamic complexity of the environment and the urgency of the system task, the preset safety threshold and / or preset collaborative obstacle avoidance range corresponding to the minimum contact distance are adjusted in real time according to preset adjustment rules.

[0127] In one embodiment, in the communication network module 101, the individual attribute data includes at least one of size, shape, maximum speed, maximum acceleration, and minimum obstacle avoidance radius;

[0128] Motion path data includes at least one of the following: target location and planned waypoints;

[0129] The motion state data of each agent includes at least one of the following: current position, velocity, acceleration, and heading angle.

[0130] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a power supply safety management method as described above.

[0131] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0132] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0133] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A multi-agent obstacle avoidance collaborative decision-making method, characterized in that, The method includes: The system acquires individual attribute data, motion path data, motion state data of each agent, as well as location data, size data, and motion data of obstacles through a pre-built communication network. Based on the acquired data, a prediction model is used to predict the future motion trajectory of each of the intelligent agents and the future state of the obstacles. Based on the future trajectory and the future state, calculate the minimum contact distance between any two agents and between each agent and an obstacle; If any of the minimum contact distances is less than a preset safety threshold, the corresponding agent and all agents within the preset cooperative obstacle avoidance range will be marked as agents that need to avoid obstacles. The communication network is used to obtain the future state of the obstacle, the individual attribute data of the intelligent agent to be avoided, and the future movement trajectory. Based on the future state, individual attribute data, and future movement trajectory, a path planning algorithm is used to generate an obstacle avoidance path. Based on the obstacle avoidance path, the individual attribute data of the intelligent agent to be avoided, and its future movement trajectory, cooperative obstacle avoidance motion parameters are calculated, and corresponding cooperative obstacle avoidance instructions are generated based on the cooperative obstacle avoidance motion parameters.

2. The method according to claim 1, characterized in that, The step of using a prediction model to predict the future motion trajectory of each intelligent agent and the future state of the obstacle includes: Based on the individual attribute data, motion path data, and motion state data of the intelligent agents, the future motion trajectory of each intelligent agent is predicted through a time-series prediction model. Based on the motion data of the obstacle, the obstacle is classified as a dynamic obstacle or a static obstacle; For the dynamic obstacle, based on its position data and motion data, the future position and motion state are predicted by a state evolution model, and the future state of the obstacle is obtained by combining the future position and motion state with the size data. For the static obstacle, the position data and size data are taken as the future state of the obstacle; The temporal prediction model includes a long short-term memory network or a temporal convolutional network, and the state evolution model includes a Kalman filter or a kinematic extrapolation model.

3. The method according to claim 2, characterized in that, The calculation of any two agents and the minimum contact distance between each agent and the obstacle includes: Based on a pre-established kinematic-geometry model, the motion state of each agent and obstacle within a preset future time interval is dynamically modeled according to the future motion trajectory of the agent and the future state of the obstacle. Based on the modeling results, the minimum Euclidean distance between any two agent size boundaries and between each agent size boundary and the obstacle size boundary is calculated in real time, and is taken as the minimum contact distance.

4. The method according to claim 1, characterized in that, The generation of obstacle avoidance paths through path planning algorithms includes: Based on the future state of the obstacle, the individual attribute data of the intelligent agent to be avoided, and the future motion trajectory, an obstacle avoidance optimization function is constructed with obstacle avoidance risk and cooperative motion cost as optimization objectives. The obstacle avoidance risk is calculated based on the dynamic relationship between the future state of the obstacle and the future motion trajectory of the intelligent agent to be avoided, and the cooperative motion cost is obtained by weighting the Euclidean distance of the velocity vectors in the future motion trajectories of each intelligent agent. Based on the improved distributed model predictive control algorithm, the obstacle avoidance optimization function is used to iteratively solve the obstacle avoidance path of each intelligent agent by satisfying the kinematic constraints limited by the individual attribute data of each agent and the preset safety threshold as the stopping conditions.

5. The method according to claim 4, characterized in that, The calculation of the cooperative obstacle avoidance motion parameters of the intelligent agent requiring obstacle avoidance, and the generation of corresponding cooperative obstacle avoidance commands based on the cooperative obstacle avoidance motion parameters, includes: Based on the obstacle avoidance risk and the cost of cooperative movement, calculate the cooperative weight coefficient of the corresponding agent; Based on the obstacle avoidance path and the cooperative weight coefficient, the motion control parameters of the corresponding intelligent agent are calculated in real time. Based on the motion control parameters, the cooperative obstacle avoidance command is generated through a preset cooperative obstacle avoidance command library.

6. The method according to claim 3, characterized in that, The method further includes: The number of obstacles is counted in real time during the dynamic modeling process; The dynamic complexity of the environment is calculated based on the number and future state of the obstacles and the future movement trajectories of each agent. Obtain the preset task time limit data of each of the aforementioned intelligent agents, and calculate the urgency of the system tasks in combination with the corresponding motion path data; Based on the dynamic complexity of the environment and the urgency of the system task, the preset safety threshold and / or preset cooperative obstacle avoidance range corresponding to the minimum contact distance are adjusted in real time according to preset adjustment rules.

7. The method according to claim 1, characterized in that, The individual attribute data includes at least one of size, shape, maximum speed, maximum acceleration, and minimum obstacle avoidance radius; The motion path data includes at least one of the target location and planned path points; The motion state data of each intelligent agent includes at least one of the following: current position, velocity, acceleration, and heading angle.

8. A multi-agent obstacle avoidance collaborative decision-making device, characterized in that, The device includes: The communication network module is used to acquire individual attribute data, motion path data, motion state data of each intelligent agent, as well as the position data, size data and motion data of obstacles through a pre-built communication network; The trajectory prediction module is used to predict the future motion trajectory of each of the intelligent agents and the future state of the obstacles based on the acquired data and a prediction model. The safe distance calculation module is used to calculate the minimum contact distance between any two intelligent agents and between each intelligent agent and an obstacle based on the future motion trajectory and the future state. The collaborative obstacle avoidance decision module is used to mark the corresponding agent and all agents within the preset collaborative obstacle avoidance range as agents that need to avoid obstacles if any of the minimum contact distances is less than a preset safety threshold. The path planning module is used to obtain the future state of the obstacle, the individual attribute data of the intelligent agent to be avoided, and the future movement trajectory through the communication network, and generate an obstacle avoidance path based on the future state, individual attribute data, and future movement trajectory using a path planning algorithm. The collaborative control module is used to calculate collaborative obstacle avoidance motion parameters based on the obstacle avoidance path, the individual attribute data of the intelligent agent to be avoided, and the future motion trajectory, and to generate corresponding collaborative obstacle avoidance commands based on the collaborative obstacle avoidance motion parameters.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images