An adaptive collision avoidance control method for a laser-guided AGV
By combining reinforcement learning models with the TEB algorithm, the linear velocity and angular velocity weights of AGVs are dynamically adjusted, solving the obstacle avoidance lag problem of laser-guided AGVs in complex dynamic environments, realizing adaptive collision avoidance control, and improving the obstacle avoidance capability and operational stability of AGVs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINGDAO JIYU AUTOMATION SYSTEM CO LTD
- Filing Date
- 2026-01-30
- Publication Date
- 2026-06-05
AI Technical Summary
Existing laser-guided AGVs struggle to effectively predict obstacle movement trends in complex and dynamic working environments, resulting in delayed obstacle avoidance responses. Furthermore, traditional methods often fail to balance safety and efficiency in dynamic environments.
By employing a reinforcement learning model combined with the TEB algorithm, vehicle pose and obstacle information are acquired through sensors, and linear velocity and angular velocity weights are dynamically adjusted to achieve adaptive collision avoidance control.
It improves the obstacle avoidance capability and operational stability of AGVs in dynamic environments, and enhances path tracking safety and efficiency while meeting kinematic constraints.
Smart Images
Figure CN122151840A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving technology, and more specifically to an adaptive collision avoidance control method for laser-guided AGVs. Background Technology
[0002] As modern industrial production continues to evolve towards automation and intelligence, traditional material handling methods relying on manual operation and semi-automation are no longer sufficient to meet the demands of today's efficient, precise, and flexible industrial production. Automated Guided Vehicles (AGVs), with their advantages of autonomous navigation, flexible scheduling, and high reliability, are widely used in warehousing management, production line material distribution, and finished product handling, becoming an important component of intelligent manufacturing systems. Based on different guidance methods, AGVs are mainly divided into three types: magnetic tape guidance, QR code guidance, and laser guidance. Among them, laser-guided AGVs, equipped with high-precision LiDAR, can achieve real-time scanning and mapping of the surrounding environment, possessing advantages such as no need for physical markers, high positioning accuracy, and strong environmental adaptability, thus occupying a dominant position in modern intelligent factories and logistics systems.
[0003] In real-world dynamic operating environments, AGVs need to safely navigate work areas containing dynamic obstacles such as personnel and forklifts while maintaining efficient task execution. Existing obstacle avoidance strategies mainly fall into two categories: global path adjustment and local path optimization. Global path adjustment avoids obstacles by recalculating the entire running path. While this can ensure safety to a certain extent, it disrupts path continuity and incurs a high computational burden, thus reducing transportation efficiency. Local path optimization methods, such as the Timed Elastic Band (TEB) algorithm, optimize the trajectory within a local area, comprehensively considering trajectory smoothness, obstacle avoidance performance, and kinematic constraints. This can improve operating efficiency while ensuring path safety. However, traditional TEB relies heavily on cost maps to describe the obstacle environment. This results in delays in updating the state of dynamic obstacles and a lack of effective prediction of obstacle movement trends, leading to lag in obstacle avoidance responses and even potential collisions in highly dynamic scenarios.
[0004] To address the aforementioned issues, some studies have attempted to combine TEB with other global planning methods or trajectory prediction techniques to improve its performance in dynamic environments. For example, one approach introduces an improved graph search algorithm in the global path planning stage to enhance path planning efficiency, and then uses TEB for local trajectory smoothing and obstacle avoidance optimization; however, this type of method often suffers from delayed response in scenarios with high obstacle speeds or irregular trajectories. Another approach introduces obstacle trajectory prediction in local path optimization, using techniques such as Kalman filtering or particle filtering to improve the estimation accuracy of obstacle motion states, and combines this with an improved dynamic window method for obstacle avoidance; however, this type of method lacks a mechanism to transform predicted information into path constraints, resulting in insufficient active obstacle avoidance capabilities, and its evaluation functions are mostly static indicators, making it difficult to balance safety and efficiency in dense dynamic environments. Reinforcement learning (RL) can continuously optimize strategies through interaction with the environment and has been widely used in the fields of mobile robot control and autonomous navigation. However, directly using reinforcement learning to generate control commands in AGVs has problems such as slow convergence speed, large training sample requirements, and difficulty in directly satisfying multiple constraints such as speed, acceleration, and obstacle avoidance, which limits its practicality in industrial scenarios.
[0005] The invention patent application with publication number CN120043527A, entitled "A Path Planning Method for Intelligent Warehouse Robots Based on TEB Dynamic Obstacle Avoidance and Improved Bidirectional A-Star Algorithm," improves global path efficiency by refining the global search strategy and achieves local obstacle avoidance by combining TEB for local trajectory optimization, thereby enhancing path continuity in static environments. However, this method only relies on the current position of obstacles in dynamic obstacle handling, lacking prediction and analysis of their movement trends. Furthermore, the weight parameters of TEB are fixed and preset, and cannot be adaptively adjusted in real time for different scenarios. Therefore, it suffers from response lag and insufficient obstacle avoidance robustness in complex dynamic environments. The invention patent application with publication number CN112835333A, entitled "A Method and System for Obstacle Avoidance and Path Planning of Multiple AGVs Based on Deep Reinforcement Learning," utilizes distributed policy learning to achieve collaborative obstacle avoidance and path optimization among multiple vehicles. Although this method has a certain degree of adaptability, its planning strategy is an end-to-end neural network output, lacking explicit guarantees for kinematic constraints such as speed and acceleration. Furthermore, it is not combined with analytical local optimizers such as TEB, making it difficult to simultaneously achieve physical feasibility and policy flexibility.
[0006] Therefore, how to effectively predict the movement trend of obstacles in complex and dynamic working environments and perform forward-looking local trajectory optimization accordingly, while achieving adaptive adjustment of optimization parameters, so as to balance obstacle avoidance safety, trajectory stability and operating efficiency under the premise of strictly meeting the kinematic constraints of AGV, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0007] In view of the above problems, the present invention proposes an adaptive collision avoidance control method for laser-guided AGVs to overcome or at least partially solve the above problems.
[0008] To achieve the above objectives, the present invention adopts the following technical solution:
[0009] This invention provides an adaptive collision avoidance control method for laser-guided AGVs, comprising: Based on the positioning and perception information obtained by the sensor components mounted on the AGV, vehicle pose information, obstacle information and global path are generated; The vehicle pose information, obstacle information, and global path are input into the trained reinforcement learning model, which outputs linear velocity weights and angular velocity weights. Based on the vehicle pose information, obstacle information, global path, linear velocity weight, and angular velocity weight, the TEB algorithm is used to optimize the local path of the AGV to obtain the optimized linear velocity and angular velocity. The AGV is controlled to move according to the optimized linear velocity and angular velocity to achieve adaptive collision avoidance.
[0010] Furthermore, the reinforcement learning model is trained through the following steps: Construct a state space, which includes longitudinal position error, heading error, minimum distance between AGV and nearest obstacle, current linear velocity, current angular velocity, rate of change of velocity, linear velocity weight, and angular velocity weight, calculated based on vehicle pose information, obstacle information, and global path. Construct a total reward function, which includes a tracking accuracy reward function, a security reward function, a stability reward function, and a task completion reward function; Based on the state space and the total reward function, the model is trained using a reinforcement learning algorithm until convergence.
[0011] Furthermore, the total reward function is expressed as:
[0012]
[0013]
[0014]
[0015]
[0016] in, For the total reward function; For tracking accuracy reward function; l 1 represents the weight of the tracking accuracy reward function; For security reward function; l 2 represents the weight of the security reward function; For stationary reward functions; l 3 represents the weights of the stationary reward function; A reward function for task completion; l 4 represents the weight of the task completion reward function; This refers to the longitudinal position error; For heading error; This represents the minimum distance between the AGV and the nearest obstacle. The safe distance between the AGV and the nearest obstacle; The rate of change of velocity; Weighted by linear velocity; The decay exponent for the linear velocity weight; Weighted by angular velocity; The decay exponent for angular velocity weighting; The severity of the penalty, among which .
[0017] Furthermore, the reinforcement learning algorithm employs the Soft Actor-Critic algorithm.
[0018] Furthermore, based on the vehicle pose information, obstacle information, global path, linear velocity weight, and angular velocity weight, the TEB algorithm is used to optimize the local path of the AGV to obtain the optimized linear velocity and angular velocity; specifically, this includes: The vehicle pose information, obstacle information, and global path are input into the optimization objective function, and the optimized AGV trajectory sequence is output. Based on the optimized trajectory sequence, the basic linear velocity and basic angular velocity are calculated; The optimized linear velocity is obtained by multiplying the base linear velocity by the linear velocity weight; and the optimized angular velocity is obtained by multiplying the base angular velocity by the angular velocity weight.
[0019] Furthermore, the optimization objective function is expressed as:
[0020] in, The overall objective function; To constrain the weights, and ; For path tracking constraints; To avoid obstacles; Linear velocity constraint; Angular velocity constraint; For kinematic constraints; For time-optimal constraints; The output trajectory of the TEB algorithm, and , x Let be the pose point of the AGV in the world coordinate system, and x =( x , y , i ), x Let x be the x-coordinate of the AGV. y The vertical coordinate of the AGV is [0, 1]. i ΔT is the heading angle of the AGV; ΔT is the time interval between adjacent pose points. Used to obtain the overall objective function Minimum trajectory B .
[0021] Furthermore, one or more constraints in the optimization objective function are implemented by penalizing the degree of constraint violation through a smoothing penalty function. Furthermore, the basic linear velocity and basic angular velocity are expressed as follows:
[0022]
[0023] in, v j For AGV in j The baseline linear velocity at any given time; oh j For AGV in j The base angular velocity at that moment; x j+1 For AGV in j The x-axis at time +1; x j For AGV in j The x-axis represents the time interval; y j+1 For AGV in j The ordinate at time +1; y j For AGV in j The vertical axis of time; for j Time and j The time interval at +1; This means normalizing the angle difference to... Within the range; i j+1 For AGV in j Heading angle at time +1; i jFor AGV in j The heading angle at any given moment.
[0024] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an adaptive collision avoidance control method for laser-guided AGVs, which has the following beneficial effects: This invention introduces reinforcement learning to dynamically weight and optimize the linear and angular velocities generated by TEB. This not only retains the advantages of TEB in satisfying kinematic and safety constraints, but also enables the AGV to adapt to the environment. Under the premise of ensuring the safe operation of the AGV, it effectively improves the system's adaptability to dynamic environments and the intelligence level of the control strategy, thereby improving the safety and stability of the AGV during operation. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0026] Figure 1 This is a schematic diagram of the adaptive collision avoidance control method for laser-guided AGV provided in an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] This invention discloses an adaptive collision avoidance control method for laser-guided AGVs, such as... Figure 1 As shown, it includes the following steps: S1. Based on the positioning and perception information obtained by the sensor components mounted on the AGV, generate vehicle pose information, obstacle information and global path; S2. Input the vehicle pose information, obstacle information and global path into the trained reinforcement learning model, and output the linear velocity weights and angular velocity weights. S3. Based on vehicle pose information, obstacle information, global path, linear velocity weight and angular velocity weight, the TEB algorithm is used to optimize the local path of the AGV to obtain the optimized linear velocity and angular velocity. S4. Control the AGV to move according to the optimized linear velocity and angular velocity to achieve adaptive collision avoidance.
[0029] This method introduces reinforcement learning to dynamically weight and optimize the linear and angular velocities generated by TEB. This not only retains the advantages of TEB in satisfying kinematic and safety constraints, but also enables AGV to adapt to the environment. Under the premise of ensuring the safe operation of AGV, it effectively improves the system's adaptability to dynamic environments and the intelligence level of control strategies, thereby improving the safety and stability of AGV during operation.
[0030] Next, each of the above steps will be explained in detail.
[0031] In step S1 above, based on the positioning and perception information obtained by the sensor components mounted on the AGV, vehicle pose information, obstacle information and global path are generated. The sensor components include a wheeled odometer, an inertial measurement unit (IMU), and a lidar. In this embodiment, the navigation stack receives pose information from the wheeled odometer, attitude information from the IMU, and point cloud data from the lidar, which together form positioning and perception information. After processing by the environmental mapping and positioning module, the vehicle pose information is output. Obstacle information, combined with task objectives and map data, is used to generate a global path through a global path planning algorithm; among which, X This represents the lateral position of the AGV. Y This represents the longitudinal position of the AGV. i The heading angle of the AGV. v This represents the current linear velocity of the AGV. oh This represents the current angular velocity of the AGV.
[0032] In step S2 above, the vehicle pose information, obstacle information and global path are input into the trained reinforcement learning model, and the linear velocity weights and angular velocity weights are output. This reinforcement learning model is trained through the following steps: (1) Construct a state space, which includes longitudinal position error, heading error, minimum distance between AGV and nearest obstacle, velocity change rate, linear velocity weight and angular velocity weight calculated based on vehicle pose information, obstacle information and global path; This state space S Represented as:
[0033] in, This refers to the longitudinal position error; For heading error; This represents the minimum distance between the AGV and the nearest obstacle.v The current linear velocity; oh The current angular velocity; The rate of change of velocity; Weighted by linear velocity; Weighted by angular velocity; (2) Construct the total reward function: The reward function for reinforcement learning is designed based on four aspects: tracking accuracy, safety, stationarity, and task completion. Tracking accuracy considers longitudinal position error and heading error; safety considers the minimum distance to obstacles; stationarity considers the rate of change of linear velocity and the rate of change of the weights of linear and angular velocities; and task completion considers whether the next local path point has been reached. The total reward function is expressed as:
[0034]
[0035]
[0036]
[0037]
[0038] in, For the total reward function; For tracking accuracy reward function; l 1 represents the weight of the tracking accuracy reward function; For security reward function; l 2 represents the weight of the security reward function; For stationary reward functions; l 3 represents the weights of the stationary reward function; A reward function for task completion; l 4 represents the weight of the task completion reward function; This refers to the longitudinal position error; For heading error; This represents the minimum distance between the AGV and the nearest obstacle. The safe distance between the AGV and the nearest obstacle; The rate of change of velocity; Weighted by linear velocity; The decay exponent for the linear velocity weight; Weighted by angular velocity; The decay exponent for angular velocity weighting; The severity of the penalty, among which .
[0039] (3) Based on the state space and the total reward function, the model is trained using a reinforcement learning algorithm until convergence. The Soft Actor-Critic (SAC) algorithm is used for reinforcement learning training to train the output linear velocity weights. With angular velocity weight In this invention, the SAC algorithm utilizes a deep neural network to approximate a high-dimensional state space, achieving continuous optimization output of the linear velocity and angular velocity weights. Compared with traditional discrete motion methods, the SAC algorithm has advantages such as smooth strategy, high training stability, fast convergence speed, and strong adaptability when dealing with continuous motion control problems. It is particularly suitable for AGV motion control scenarios that require real-time dynamic adjustment of speed strategies, as described in this invention.
[0040] (4) Empirical sample extraction: After the AGV completes a trajectory execution and collects the corresponding environmental interaction data, it can enter the reinforcement learning training phase. The Actor network and Critic network in the SAC algorithm update their network parameters to optimize the strategy; the Actor network is updated based on the action state value output by the Critic network, while the Critic network is updated based on the reward fed back by the environment after selecting an action; the update requires randomly drawing a batch of experience samples from the experience pool and distributing them to each network.
[0041] (5) Experience replay: The quadruplets given in the randomly selected empirical sample Break it down into its components to form the current state. Current action Current Rewards and the next state The vectors are transmitted into the Actor and Critic networks for experience replay; the loss function of each network is calculated using quadruples, and the parameters are updated using gradient descent backpropagation, thereby learning the updated safety control strategy and better completing the motion control task.
[0042] The loss function of the Actor network is expressed as:
[0043] in, Let be the loss function of the Actor network; This is the expected value; The weights for entropy; For strategy In the current state Select the current action below The logarithmic probability; In the current state Take the current action Expected returns; Estimating the value function for the Critic network.
[0044] The loss function of the Critic network is expressed as:
[0045] in, is the loss function of the Critic network; This is a discount factor used to balance the importance of immediate rewards and future rewards; For the next state Next action Expected returns; To select the minimum Q value from the Critic network and reduce the variance of the estimate; For strategy In the next state Select the next action The logarithmic probability.
[0046] Based on the two loss functions mentioned above, the Actor network and Critic network are continuously updated until the policy converges stably, thereby achieving effective learning of the linear velocity and angular velocity weights of the AGV.
[0047] In step S3 above, based on vehicle pose information, obstacle information, global path, linear velocity weight, and angular velocity weight, the TEB algorithm is used to optimize the local path of the AGV, obtaining the optimized linear velocity and angular velocity; specifically including: S31. Input the vehicle pose information, obstacle information, and global path into the optimization objective function, and output the optimized AGV trajectory sequence; specifically: (1) Trajectory parametric modeling: The mathematical model of the TEB time elastic band algorithm consists of the pose sequence of the AGV in the world coordinate system, expressed as:
[0048] in, For pose sequence; For a series of pose points of the AGV trajectory, there are a total of n One point.
[0049] The TEB algorithm, in addition to considering pose information, also considers the time information of the trajectory, as follows:
[0050] in, T It is a time series; The time interval between adjacent poses, i.e. j Time and j The time interval at +1 is in total n -1 time period, and ; By combining the pose sequence and the time sequence described above, we can obtain the time-based trajectory sequence of the TEB algorithm, represented as:
[0051] in, This is the pose sequence of the AGV in the world coordinate system, which contains a total of n One pose point (excluding the initial pose point); ( x 0, y 0, i 0) represents the AGV at the starting point ( x 0, y The pose point of 0); x j , y j , i j ) for AGV in j The pose point at any given moment; x j For AGV in j The x-axis represents the time interval; y j For AGV in j The vertical axis of time; i j For AGV in j The heading angle at any given moment; for j Time and j The time interval at +1; (2) Construction of objective function and setting of constraints: Considering six constraints—path tracking, obstacle avoidance, linear velocity, angular velocity, kinematics, and time optimization—the optimization objective function for TEB is established as follows:
[0052] in, The overall objective function; To constrain the weights, and ; For path tracking constraints; To avoid obstacles; Linear velocity constraint; Angular velocity constraint; For kinematic constraints; For time-optimal constraints; The output trajectory of the TEB algorithm, and , x Let be the pose point of the AGV in the world coordinate system, and x =( x , y , i ), x Let x be the x-coordinate of the AGV. y The vertical coordinate of the AGV is [0, 1]. i ΔT is the heading angle of the AGV; ΔT is the time interval between adjacent pose points. Used to obtain the overall objective function Minimum trajectory B .
[0053] (3) Soft constraint transformation: The constraints in the aforementioned objective function are called hard constraints, which are conditions that must be satisfied during the optimization process. However, the existence of hard constraints can lead to unsolvable optimization problems or excessively long solution times. In AGV trajectory planning, this often results in untimely AGV control, leading to loss of control and safety issues. Therefore, the TEB algorithm transforms the hard constraints into soft constraints and adds them to the objective function in the form of a cost function. The cost function in TEB is defined as a piecewise, continuously differentiable function, expressed as:
[0054] in, A smooth penalty function for the degree of constraint violation; These are the observations for the current objective problem; The constraint weights for the current problem; This is the boundary offset factor; k The factor that amplifies the cost function; l It is the order.
[0055] (4) Cost function setting: After completing the constraint softening process, based on the above optimization objective function, specific settings are made for each cost function to achieve effective constraints on the trajectory optimization process, expressed as follows:
[0056] in, For path tracking constraints; The distance from the trajectory point to the reference path point; This represents the maximum non-penalty boundary where a pathpoint deviates from the global pathpoint.
[0057]
[0058] in, To avoid obstacles; The distance from the trajectory point to the obstacle; This is the minimum permissible path point distance from the obstacle avoidance safety distance.
[0059]
[0060] in, Linear velocity constraint; v The current linear velocity; This represents the maximum linear velocity.
[0061]
[0062] in, Angular velocity constraint; oh The current angular velocity; This represents the maximum angular velocity.
[0063]
[0064] in, Nonholonomic kinematic constraints; x j+1 For AGV in j The x-axis at time +1; x j For AGV in j The x-axis represents the time interval; The Euclidean distance between two adjacent pose points; i j+1 For AGV in j Heading angle at time +1; i j For AGV in j The heading angle at any given moment.
[0065]
[0066] in, For time-optimal constraints; n The number of pose points.
[0067] (5) Trajectory solving based on graph optimization: The TEB algorithm utilizes graph optimization to transform the sparse optimization problem into a hypergraph. A hypergraph is a graph where an edge connects multiple nodes. In TEB, the nodes of the hypergraph represent the robot's spatial pose and time interval sequence, while the hyperedges represent the aforementioned constraints. Finally, the output trajectory of TEB is obtained using the graph optimization framework "g2o framework". .
[0068] S32. Based on the optimized trajectory sequence, calculate the basic linear velocity and basic angular velocity; the linear velocity and angular velocity of this invention are calculated based on the pose change and time interval, and are expressed as follows:
[0069]
[0070] in, v j For AGV in j The baseline linear velocity at any given time; oh j For AGV in j The base angular velocity at that moment; x j+1 For AGV in j The x-axis at time +1; x j For AGV in j The x-axis represents the time interval; y j+1 For AGV in j The ordinate at time +1; y j For AGV in j The vertical axis of time; for j Time and j The time interval at +1; This means normalizing the angle difference to... Within the range; i j+1 For AGV in j Heading angle at time +1; i j For AGV in j The heading angle at any given moment.
[0071] S33. Multiply the basic linear velocity by the linear velocity weight to obtain the optimized linear velocity; and multiply the basic angular velocity by the angular velocity weight to obtain the optimized angular velocity.
[0072] In step S4 above, the AGV is controlled to move according to the optimized linear velocity and angular velocity to achieve adaptive collision avoidance.
[0073] In summary, the adaptive collision avoidance control method for laser-guided AGVs provided by this invention constructs a collaborative control system consisting of a navigation stack, reinforcement learning, TEB (Tracking Equipment), and the AGV, achieving adaptive motion control of the AGV. This method fully combines the advantages of TEB local trajectory optimization and reinforcement learning to construct an adaptive collision avoidance control strategy for laser-guided AGVs. Specifically, TEB utilizes analytical optimization methods to simultaneously consider multiple constraints during trajectory generation, ensuring trajectory smoothness and physical feasibility. Reinforcement learning, through its autonomous learning capabilities, dynamically outputs the weight coefficients of linear velocity and angular velocity in TEB, achieving online adaptive adjustment of the linear velocity and angular velocity steering strategy, thereby realizing adaptive obstacle avoidance under different operating conditions. Through the deep integration of reinforcement learning and TEB, this method forms a closed-loop linkage between trajectory optimization and dynamic weight adjustment. It not only retains the advantages of TEB in satisfying kinematic and safety constraints but also introduces data-driven environmental adaptability, achieving comprehensive optimization of multiple objectives such as safety, stability, and efficiency, significantly improving the obstacle avoidance capability and path tracking stability of the AGV in complex dynamic environments. Furthermore, this method can be directly embedded into the existing AGV control architecture, reducing the difficulty of engineering deployment and showing broad application prospects in the field of AGV autonomous navigation and intelligent logistics scheduling.
[0074] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0075] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An adaptive collision avoidance control method for laser-guided AGVs, characterized in that, include: Based on the positioning and perception information obtained by the sensor components mounted on the AGV, vehicle pose information, obstacle information and global path are generated; The vehicle pose information, obstacle information, and global path are input into the trained reinforcement learning model, which outputs linear velocity weights and angular velocity weights. Based on the vehicle pose information, obstacle information, global path, linear velocity weight, and angular velocity weight, the TEB algorithm is used to optimize the local path of the AGV to obtain the optimized linear velocity and angular velocity. The AGV is controlled to move according to the optimized linear velocity and angular velocity to achieve adaptive collision avoidance.
2. The adaptive collision avoidance control method for laser-guided AGV as described in claim 1, characterized in that, The reinforcement learning model is trained through the following steps: Construct a state space, which includes longitudinal position error, heading error, minimum distance between AGV and nearest obstacle, current linear velocity, current angular velocity, rate of change of velocity, linear velocity weight, and angular velocity weight, calculated based on vehicle pose information, obstacle information, and global path. Construct a total reward function, which includes a tracking accuracy reward function, a security reward function, a stability reward function, and a task completion reward function; Based on the state space and the total reward function, the model is trained using a reinforcement learning algorithm until convergence.
3. The adaptive collision avoidance control method for laser-guided AGVs as described in claim 2, characterized in that, The total reward function is expressed as follows: in, For the total reward function; For tracking accuracy reward function; λ 1 represents the weight of the tracking accuracy reward function; For security reward function; λ 2 represents the weight of the security reward function; For stationary reward functions; λ 3 represents the weights of the stationary reward function; A reward function for task completion; λ 4 represents the weight of the task completion reward function; This refers to the longitudinal position error; For heading error; This represents the minimum distance between the AGV and the nearest obstacle. The safe distance between the AGV and the nearest obstacle; The rate of change of velocity; Weighted by linear velocity; The decay exponent for the linear velocity weight; Weighted by angular velocity; The decay exponent for angular velocity weighting; The severity of the penalty, among which .
4. The adaptive collision avoidance control method for laser-guided AGV as described in claim 2, characterized in that, The reinforcement learning algorithm used is the Soft Actor-Critic algorithm.
5. The adaptive collision avoidance control method for laser-guided AGV as described in claim 1, characterized in that, Based on the vehicle pose information, obstacle information, global path, linear velocity weight, and angular velocity weight, the TEB algorithm is used to optimize the local path of the AGV to obtain the optimized linear velocity and angular velocity; specifically including: The vehicle pose information, obstacle information, and global path are input into the optimization objective function, and the optimized AGV trajectory sequence is output. Based on the optimized trajectory sequence, the basic linear velocity and basic angular velocity are calculated; The optimized linear velocity is obtained by multiplying the base linear velocity by the linear velocity weight; and the optimized angular velocity is obtained by multiplying the base angular velocity by the angular velocity weight.
6. The adaptive collision avoidance control method for laser-guided AGV as described in claim 5, characterized in that, The optimization objective function is expressed as: in, The overall objective function; To constrain the weights, and ; For path tracking constraints; To avoid obstacles; Linear velocity constraint; Angular velocity constraint; For kinematic constraints; For time-optimal constraints; The output trajectory of the TEB algorithm, and , ξ Let be the pose point of the AGV in the world coordinate system, and ξ =( x , y , θ ), x Let x be the x-coordinate of the AGV. y The vertical coordinate of the AGV is [0, 1]. θ ΔT is the heading angle of the AGV; ΔT is the time interval between adjacent pose points. Used to obtain the overall objective function Minimum trajectory B .
7. The adaptive collision avoidance control method for laser-guided AGV as described in claim 6, characterized in that, One or more constraints in the optimization objective function are achieved by penalizing the degree of constraint violation through a smoothing penalty function.
8. The adaptive collision avoidance control method for laser-guided AGV as described in claim 5, characterized in that, The basic linear velocity and basic angular velocity are respectively expressed as: in, v j For AGV in j The baseline linear velocity at any given time; ω j For AGV in j The base angular velocity at that moment; x j+1 For AGV in j The x-axis at time +1; x j For AGV in j The x-axis represents the time interval; y j+1 For AGV in j The ordinate at time +1; y j For AGV in j The vertical axis of time; for j Time and j The time interval at +1; This means normalizing the angle difference to... Within the range; θ j+1 For AGV in j Heading angle at time +1; θ j For AGV in j The heading angle at any given moment.