Limited space large equipment transportation system and method based on improved A* algorithm and deep reinforcement learning

Abstract

The invention discloses a limited space large-scale equipment transportation system and method based on an improved A * algorithm and deep reinforcement learning, and the system comprises an initial planning module, a global path optimization module and a local path dynamic optimization module. The global path optimization module is used for generating a global optimal transport path and planning an initial global transport path of the large-scale equipment based on an A * algorithm, the global path optimization module is used for adjusting weight coefficients of actual cost and heuristic cost in a heuristic function based on the initial global transport path to generate the global optimal transport path, and the local path dynamic optimization module is used for dynamically optimizing the global optimal transport path based on a deep reinforcement learning algorithm. A local planning network of an Actor-Critic architecture is constructed, obstacle information is acquired in real time, postures of a local transportation path and the large equipment are dynamically adjusted based on the obstacle information and sliding window information of a global optimal transportation path, and a collision-free optimal transportation path is generated, so that the transportation efficiency and safety of the large equipment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large equipment transportation, and particularly relates to a large equipment transportation system in a limited space based on an improved A* algorithm and deep reinforcement learning. BACKGROUND

[0002] In the transportation process of large equipment, various challenges are often faced, such as the interior of a high-rise building on a construction site. Due to the limited operation space, traditional transportation methods such as cranes and other mechanical equipment may not be used, so that the transfer operation of large components cannot be conveniently and quickly performed, the equipment is inconvenient to be in place, and the turning angle through which the equipment can pass cannot be accurately judged.

[0003] The A* algorithm combines the actual path cost and the estimated path cost, efficiently finds the optimal path from the starting point to the ending point, is widely used in the field of path planning, and can solve the shortest path for the transportation of large equipment.

[0004] A path planning method based on an improved A* algorithm is disclosed in Chinese Patent No. CN 111780777 A, which comprises: establishing an initialized grid cost map according to environment information; planning a global path by using an improved A* algorithm; based on the global path and the performance of a laser radar sensor, designing a sliding window, taking the information detected by the window as the state input of the network; based on a deep reinforcement learning method, adopting an Actor-Critic architecture, designing a local planning network, and combining knowledge and data methods, the optimal path can be quickly planned.

[0005] However, the existing planning system based on the A* algorithm has low search efficiency, long planning time and weak adaptability, and it is difficult to meet the efficient and safe needs of large equipment transportation in a limited space.

[0006] Therefore, how to improve an efficient and safe large equipment transportation system has become a problem to be solved in the field. SUMMARY

[0007] In view of the defects of the prior art, the purpose of the present application is to provide an efficient and safe large equipment transportation system and method in a limited space based on an improved A* algorithm and deep reinforcement learning.

[0008] In order to achieve the above purpose, the large equipment transportation system in a limited space based on the improved A* algorithm and deep reinforcement learning provided by the present application comprises an initial planning module, a global path optimization module and a local path dynamic optimization module, The initial planning module can extract transportation environment information, construct an initial grid map, and plan an initial global transportation path for the large equipment based on an A* algorithm. The global path optimization module can adjust the weight coefficients of the actual cost and heuristic cost in the heuristic function based on the initial global transportation path, generate a globally optimal transportation path, and the local path dynamic optimization module can construct a local planning network based on an Actor-Critic architecture and real-time obstacle information based on a deep reinforcement learning algorithm. The local path dynamic optimization module can dynamically adjust the local transportation path and the pose of the large equipment based on the sliding window information of the global optimal transportation path and the obstacle information, and generate a collision-free optimal transportation path.

[0009] Further, the initial planning module extracts obstacle positions, large equipment sizes, and spatial restriction information of the transportation environment based on positioning and map construction techniques. After constructing an initial grid map based on transportation environment information, the initial planning module establishes basic data for path planning, including the starting point, end point, obstacle positions, and large equipment model of the transportation.

[0010] Further, the global path optimization module determines the weight coefficients of the actual cost and heuristic cost based on the number of nodes and road conditions of the initial global transportation path, and optimizes the initial global transportation path.

[0011] Further, the global path optimization module establishes an open set and a closed set, adds the starting point to the open set as the highest priority, and performs main loop calculation. When the open set is not empty, the node with the highest priority is selected from the open set, and end point judgment is performed. If the node is the end point, path backtracking is performed to return to the structure. If the node is not the end point, the node is removed from the open set to the closed set, and all adjacent nodes of the node are traversed.

[0012] Further, for each adjacent node, if the adjacent node is already in the closed set, the next adjacent node is skipped. If the adjacent node is not in the open set, the parent node of the adjacent node is set and priority calculation is performed, and the adjacent node is added to the open set for looping.

[0013] Further, the local path dynamic optimization module includes an Actor network and a Critic network. The Actor network interacts with environment information, large equipment pose, sliding window information, and obstacle information, and outputs policy parameters to generate corresponding actions based on the current state. The Critic network calculates the Q value of the action generated by the Actor network based on the current state, and updates the policy parameters of the Actor network through backpropagation. Meanwhile, the Critic network updates its own Critic network parameters by minimizing the error between the predicted Q value and the actual TD target value.

[0014] Furthermore, the Actor network and the Critic network are each split into an online network and a target network. The online network is updated with the latest network parameters at each time step to achieve real-time learning. The parameters of the target network are periodically copied from the online network and are not updated; they are only used to calculate the TD target value.

[0015] Furthermore, during the Critic network update, the action for the next state is calculated through the target Actor network, and the target TD value for the next state is calculated through the target Critic network. Then, the gradient descent method is used to minimize the learning rate to evaluate the error between the predicted Q value and the actual Q value.

[0016] Furthermore, when the Actor network is updated, the policy gradient is determined by calculating the loss function of the Actor network.

[0017] To achieve the above objectives, the present invention provides a method for transporting large equipment in a confined space based on an improved A* algorithm and deep reinforcement learning, comprising: The initial planning module extracts transportation environment information and establishes an initial grid map, then establishes basic data for path planning, and plans the initial global transportation path for large equipment based on the A* algorithm; The global path optimization module determines the weight coefficients of the actual cost and the heuristic cost based on the initial global transportation path optimization heuristic function, and generates the globally optimal transportation path through the main loop calculation of nodes in the open set and the closed set. Based on the globally optimal transportation path, information about the path ahead is detected through a sliding window. The local path dynamic optimization module is based on deep reinforcement learning algorithm. It establishes a local planning network with an Actor-Critic architecture through Actor network and Critic network. Based on obstacle information and sliding window information of the global optimal transportation path, it dynamically adjusts the local transportation path and the attitude of large equipment to generate a collision-free optimal transportation path.

[0018] The present invention provides a confined space large equipment transportation system and method based on an improved A* algorithm and deep reinforcement learning. First, an initial global transportation path is planned using the A* algorithm. Then, the weight coefficients of the actual cost and heuristic cost are adjusted to optimize the global transportation path, which can improve the efficiency of path planning. Next, the local transportation path and the attitude of the large equipment are dynamically adjusted by combining obstacle information and sliding window information of the global optimal transportation path. Therefore, by combining the optimization of global and local transportation paths, a collision-free optimal path can be generated to ensure the safety and efficiency of large equipment transportation. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0020] Figure 1 A block diagram of a confined space large equipment transportation system based on an improved A* algorithm and deep reinforcement learning provided by the present invention; Figure 2 This is a schematic diagram illustrating the operation of the initial planning module and the global path optimization module in this invention; Figure 3 This is a schematic diagram illustrating the operation of the local path dynamic optimization module in this invention; Figure 4 This is a flowchart of the method for transporting large equipment in a confined space based on the improved A* algorithm and deep reinforcement learning in this invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0022] See Figure 1 The image shows an example of a large-scale equipment transportation system in a confined space based on an improved A* algorithm and deep reinforcement learning, provided by the present invention.

[0023] As shown in the figure, the large equipment transportation system in the confined space based on the improved A* algorithm and deep reinforcement learning in this example mainly includes the initial planning module 1, the global path optimization module 2, and the local path dynamic optimization module 3.

[0024] The initial planning module 1 can extract transportation environment information, construct an initial grid map, and plan the initial global transportation path of large equipment based on the A* algorithm. The global path optimization module 2 can adjust the weight coefficients of the actual cost and heuristic cost in the heuristic function based on the initial global transportation path to generate the global optimal transportation path. The local path dynamic optimization module 3 can construct a local planning network with an Actor-Critic architecture based on a deep reinforcement learning algorithm, acquire obstacle information in real time, and dynamically adjust the local transportation path and the attitude of the large equipment based on the obstacle information and the sliding window information of the global optimal transportation path to generate a collision-free optimal transportation path, so as to ensure the safety and efficiency of the transportation of large equipment.

[0025] The initial planning module 1 first extracts information on the location of obstacles, the size of large equipment, and spatial constraints in the transportation environment based on positioning and mapping technology to obtain transportation environment information. Then, it constructs an initial grid map based on the transportation environment information to provide basic data for the transportation path planning of large equipment.

[0026] Furthermore, the initial planning module 1 establishes the basic data for path planning, which includes the starting point, ending point, obstacle locations, and large equipment models of the transportation, enabling the initial planning module to perform initial global transportation path planning for large equipment using the A* algorithm based on the basic data.

[0027] Combination Figure 2 The specific calculations for initial global transportation path planning based on the A* algorithm are as follows: 1) Initialization: The initial planning module 1 adds the starting point S to the open list and sets the actual cost g to 0. The evaluation function F is equal to the heuristic cost h, i.e., F=g.

[0028] 2) Loop check: Check if the open list is empty. If it is empty, the search ends; otherwise, continue to the next step to select nodes.

[0029] 3) Node selection: Select the node N with the smallest F value from the open list.

[0030] 4) Target determination: If node N is the target point, then construct the path and end the search.

[0031] 5) Node processing: Move node N into the closed list, indicating that it has been processed.

[0032] 6) Adjacency Node Determination: Traverse each adjacent node of node N and perform the following determination: a. If the node is not accessible or is already in the closed list, return to step 2); b. Otherwise, proceed to the next step.

[0033] 7) Open list update: a. If the node is not in the open list, add it to the open list, update its F value, and then return to step 2); b. Otherwise, proceed to the next step.

[0034] 8) Cost Comparison: Calculate the g and F values ​​of the node and compare them with the old values: a. If the new g value is not less than the old g value, then return to step 2); b. Otherwise, proceed to the next step.

[0035] 9) Update information: Update the g and F values ​​of node, set the parent node of node to N, and then return to step 2).

[0036] Therefore, the initial planning module 1 performs node cyclic judgment calculation based on the A* algorithm. When the destination is added to the open list and becomes the current node, the initial global transportation path of the large equipment is generated.

[0037] To improve the search efficiency of the initial planning module 1 and narrow the search range, the global path optimization module 2 can modify the A* algorithm based on the initial global transportation path by optimizing the heuristic function, adjusting the weight coefficients of the actual cost and the heuristic cost in the heuristic function, and generating the globally optimal transportation path.

[0038] Combination Figure 2 Specifically, the global path optimization module 2 uses the following formula to adjust the heuristic function. ; Where value1 is the weight coefficient of the actual cost g, and value2 is the weight coefficient of the heuristic cost h. When value1=1 and value2=1, it is the A* algorithm of the current initial planning module 1. When value2 increases, the algorithm will pay more attention to heuristic information, that is, it will be more inclined to search in the direction of the destination, thereby improving the search and computation efficiency. However, when value2 is too large, the heuristic function may no longer satisfy the tolerance, thus failing to guarantee finding the optimal path.

[0039] Therefore, in practical applications, if a more accurate path (i.e. closer to the shortest path) is needed, value1 can be increased appropriately to make the algorithm pay more attention to the actual path length already traversed. If a faster running speed is needed, value2 can be increased appropriately to improve search efficiency.

[0040] To balance the efficiency and accuracy of path planning, the global path optimization module 2 determines and adjusts the weight coefficients value1 and value2 of the actual cost and heuristic cost based on the number of nodes and road conditions in the initial global transportation path, so as to improve the overall running efficiency of the algorithm and the accuracy of the path.

[0041] As an example, value2 is usually between 1 and 5, and can be adjusted adaptively according to the specific application.

[0042] Furthermore, the global path optimization module 2 optimizes the A* algorithm of the initial planning module 1 based on the heuristic function, and generates the globally optimal transportation path based on the optimized A* algorithm.

[0043] Specifically, the calculation of the globally optimal transportation path based on the optimized A* algorithm is as follows: 1) Initialization: Create open sets (open_set) and close sets (close_set).

[0044] 2) Starting point setting: Add the starting point to the open set and assign it a priority of 0, which indicates the highest priority.

[0045] 3) Main loop: When the open set is not empty, execute the following steps: a. Selecting a node: Select the node n with the highest priority from the open set.

[0046] b. Endpoint determination: If node n is the endpoint, then execute the following sub-steps: a) Path backtracking: Starting from the end point, backtrack to the starting point through the parent pointer.

[0047] b) Return result: Output the found path, and the algorithm terminates.

[0048] c. Non-terminal processing: If node n is not a terminal, then perform the following sub-steps: a) Set transfer: Remove node n from the open set and add it to the closed set.

[0049] b) Adjacency node traversal: Traverse all adjacent nodes of node n. For each adjacent node m, perform the following judgment: i. Closed set check: If node m is already in the closed set, skip it and continue to the next adjacent node.

[0050] ii. Open set check: If node m is not in the open set, perform the following operation: Set the parent node: Set the parent pointer of node m to point to node n.

[0051] Priority Calculation: Calculate the priority of node m using the following formula: g(m) = g(n) + cost(n, m); h(m) = distance(m, destination); f(m) = value1 * g(m) + value2 * h(m).

[0052] Add to open set: Add node m to the open set.

[0053] Therefore, the global path optimization module 2 performs node cyclic judgment calculations based on the optimized A* algorithm. When the destination is added to the open set and becomes the current node, the globally optimal transportation path for large equipment is generated. At the same time, it can improve the search efficiency of the A* algorithm, shorten the planning time, and reduce the search space.

[0054] Furthermore, based on the globally optimal transportation path, this transportation system uses a sliding window to detect the path information ahead in order to obtain local path information in front of large equipment, such as a local segment consisting of N nodes before and after the current moment in the globally optimal path.

[0055] To ensure that large equipment can interact with the transportation environment in real time while being transported along the globally optimal transportation path, dynamically adjust the transportation path and the attitude of the large equipment, prevent collisions between the large equipment and obstacles, and ensure transportation efficiency and safety, this transportation system also includes a local path dynamic optimization module 3. The local path dynamic optimization module 3 can construct a local planning network with an Actor-Critic architecture based on a deep reinforcement learning algorithm, and acquire obstacle information in real time. Based on the obstacle information and the sliding window information of the globally optimal transportation path, it dynamically adjusts the local transportation path and the attitude of the large equipment to generate a collision-free optimal transportation path.

[0056] Specifically, the local path dynamic optimization module 3 uses an environmental detection device, such as a lidar, to acquire obstacle information of the global optimal transportation path in real time, including obstacle location, size, and movement trend.

[0057] Furthermore, the local path dynamic optimization module 3 combines the heterogeneous policy Actor-Critic framework with deep learning techniques, including Actor networks and Critic networks.

[0058] Combination Figure 3 The Actor network interacts with environmental information, the attitude of large equipment, sliding window information, and obstacle information, and outputs policy parameters based on the current state. This generates corresponding actions for the transportation system, such as the directional offset of local paths, the adjustment of the attitude angle of large equipment, and the correction of transportation speed.

[0059] The Critic network can calculate the Q-value of an action based on the action generated by the Actor network and the current state. This Q-value represents the expected cumulative reward after performing the action in the current state, and it provides the basis for updating the Actor network. The Q-value is used as the weight of the Actor network's policy gradient, and backpropagation is used to update the Actor network's policy parameters. .

[0060] Meanwhile, the Critic network can also continuously iterate itself, updating its own Critic network parameters w by minimizing the error between the predicted Q value and the actual Q value (approximately the TD target value), thereby improving the accuracy of the Actor network evaluation and providing a reliable basis for updating the Actor network.

[0061] To improve the reliability of the local path dynamic optimization module 3, the Actor network and Critic network are split into two sub-networks, the online network and the target network, respectively, by setting different network parameters while keeping the neural network structure unchanged.

[0062] In this system, the online network is updated with the latest network parameters at each time step to achieve real-time learning. The parameters of the target network are periodically copied from the online network and are not updated; they are only used to calculate the TD target value. The difference between the parameters of the online network and the target network severs the correlation between them. Therefore, by periodically updating the parameters of the target network, the TD target value can be made more stable, thereby improving the training efficiency and stability of the Local Path Dynamic Optimization Module 3.

[0063] As an example, the local path dynamic optimization module 3 constructs an Actor-Critic network architecture, including an online Actor network, a target Actor network, an online Critic network, and a target Critic network. Therefore, the online Actor network takes the current state as its input. It outputs the action; the online Critic network inputs the current state. The target Actor network has the same structure as the online Actor mesh, with initial parameters copied from the online Actor mesh, and is used to compute the next state. The optimal action is obtained by copying the initial parameters from the online Critic network to the target network. The target network has the same structure as the online Critic network and is used to calculate a stable TD target value. In this way, the online network is responsible for real-time learning and the target network is responsible for stable evaluation. This can effectively solve the problem of TD target value oscillation during single network training and improve the training efficiency and stability of the local path dynamic optimization module 3.

[0064] Furthermore, during the transportation of large equipment along the globally optimal transport path, the Actor network interacts with environmental information, the large equipment's attitude, sliding window information, and obstacle information. Based on the current state, it outputs and executes actions, and then returns a reward. and the next state This generates a large number of experience samples of "state-action-reward-next state" and stores them in the experience pool.

[0065] Next, a batch of samples is randomly sampled from the experience pool, and the next state is calculated through the target Actor network. action The next state is calculated using the target Critic network. The TD target value is calculated as follows: .

[0066] Furthermore, the Critic network is updated using gradient descent to minimize the error between the evaluated Q-value and the actual Q-value, thereby minimizing the error between the predicted Q-value and the TD target value. The specific calculation is as follows:

[0067] Next, the Actor network is updated, and the policy gradient is calculated using the loss function of the Actor network. The specific calculation is as follows:

[0068] The gradient ascent method is used to update the Actor network parameters to optimize the Actor network's decision-making strategy, enabling the Actor network to generate better actions in the next state. This optimizes the directional offset of the local path, the adjustment of the large equipment's attitude angle, and the correction of the transportation speed, thereby generating a collision-free optimal transportation path.

[0069] This constitutes the large-scale equipment transportation system in a confined space based on the improved A* algorithm and deep reinforcement learning provided by the present invention.

[0070] This invention also provides a transportation method for a confined space large equipment transportation system based on the above-described scheme, using an improved A* algorithm and deep reinforcement learning. This transportation method includes: Combination Figure 4 First, the initial planning module 1 extracts transportation environment information and establishes an initial grid map, then establishes the basic data for path planning, and plans the initial global transportation path for large equipment based on the A* algorithm.

[0071] Next, the global path optimization module 2 determines the weight coefficients of the actual cost and the heuristic cost based on the initial global transportation path optimization heuristic function. Based on the optimized A* algorithm, it generates the globally optimal transportation path through the main loop calculation of nodes in the open set and the closed set.

[0072] Furthermore, based on the globally optimal transportation path, information about the path ahead is detected through a sliding window.

[0073] At this point, the local path dynamic optimization module 3, based on the deep reinforcement learning algorithm, establishes a local planning network with an Actor-Critic architecture through the Actor network and Critic network. Based on obstacle information and the sliding window information of the global optimal transportation path, it dynamically adjusts the local transportation path and the attitude of the large equipment to generate a collision-free optimal transportation path.

[0074] The present invention provides a confined space large equipment transportation system and method based on an improved A* algorithm and deep reinforcement learning. First, an initial global transportation path is planned using the A* algorithm. Then, the weight coefficients of the actual cost and heuristic cost are adjusted to optimize the global transportation path, which can improve the efficiency of path planning. Next, the local transportation path and the attitude of the large equipment are dynamically adjusted by combining obstacle information and sliding window information of the global optimal transportation path. Therefore, by combining the optimization of global and local transportation paths, a collision-free optimal path can be generated to ensure the safety and efficiency of large equipment transportation.

[0075] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A confined space large equipment transportation system based on an improved A* algorithm and deep reinforcement learning, characterized in that, It includes an initial planning module, a global path optimization module, and a local path dynamic optimization module. The initial planning module can extract transportation environment information, construct an initial grid map, and plan the initial global transportation path of large equipment based on the A* algorithm. The global path optimization module can adjust the weight coefficients of the actual cost and the heuristic cost in the heuristic function based on the initial global transportation path to generate the global optimal transportation path. The local path dynamic optimization module can construct a local planning network with an Actor-Critic architecture based on a deep reinforcement learning algorithm, acquire obstacle information in real time, and dynamically adjust the local transportation path and the attitude of the large equipment based on the obstacle information and the sliding window information of the global optimal transportation path to generate a collision-free optimal transportation path.

2. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning according to claim 1, characterized in that, The initial planning module extracts obstacle locations, large equipment dimensions, and space constraint information from the transportation environment based on positioning and mapping technology. After constructing an initial grid map based on the transportation environment information, it establishes basic data for path planning. The basic data includes the starting point, ending point, obstacle locations, and large equipment models of the transportation.

3. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning according to claim 2, characterized in that, The global path optimization module determines the weight coefficients of the actual cost and the heuristic cost based on the number of nodes and road conditions of the initial global transportation path, and optimizes the initial global transportation path.

4. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning according to claim 3, characterized in that, The global path optimization module establishes an open set and a closed set. The starting point is added to the open set with the highest priority and the main loop is used for calculation. When the open set is not empty, the node with the highest priority is selected from the open set and the destination is determined. If the node is the destination, the path backtracking is performed to return to the structure. If the node is not the destination, the node is removed from the open set and moved to the closed set, and all adjacent nodes of the node are traversed.

5. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning according to claim 4, characterized in that, For each adjacent node, if the adjacent node is already in the closed set, skip to the next adjacent node; if the adjacent node is not in the open set, set the parent node of the adjacent node, calculate the priority, and add the adjacent node to the open set for looping.

6. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning according to claim 1, characterized in that, The local path dynamic optimization module includes an Actor network and a Critic network. The Actor network interacts with environmental information, the attitude of large equipment, sliding window information, and obstacle information, and outputs policy parameters based on the current state to generate corresponding actions. The Critic network calculates the Q-value of the action based on the action generated by the Actor network and the current state, and updates the policy parameters of the Actor network through backpropagation. At the same time, the Critic network updates its own Critic network parameters by minimizing the error between the predicted Q-value and the actual TD target value.

7. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning as described in claim 6, characterized in that, The Actor network and the Critic network are each split into an online network and a target network. The online network is updated with the latest network parameters at each time step to achieve real-time learning. The parameters of the target network are periodically copied from the online network and are not updated; they are only used to calculate the TD target value.

8. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning as described in claim 7, characterized in that, For updating the Critic network, the action of the next state is calculated through the target Actor network, and the TD target value of the next state is calculated through the target Critic network. Then, the gradient descent method is used to minimize the learning rate to evaluate the error between the predicted Q value and the actual Q value.

9. The confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning as described in claim 8, characterized in that, For updating the Actor network, the policy gradient is determined by calculating the loss function of the Actor network.

10. A transportation method for a confined space large equipment transportation system based on the improved A* algorithm and deep reinforcement learning according to any one of claims 1 to 9, characterized in that, The transportation method includes: The initial planning module extracts transportation environment information and establishes an initial grid map, then establishes basic data for path planning, and plans the initial global transportation path for large equipment based on the A* algorithm; The global path optimization module determines the weight coefficients of the actual cost and the heuristic cost based on the initial global transportation path optimization heuristic function, and generates the globally optimal transportation path through the main loop calculation of nodes in the open set and the closed set. Based on the globally optimal transportation path, information about the path ahead is detected through a sliding window. The local path dynamic optimization module is based on deep reinforcement learning algorithm. It establishes a local planning network with an Actor-Critic architecture through Actor network and Critic network. Based on obstacle information and sliding window information of the global optimal transportation path, it dynamically adjusts the local transportation path and the attitude of large equipment to generate a collision-free optimal transportation path.

Images