Rapid task planning method based on deep learning

Abstract

The invention discloses a rapid task planning method based on deep learning. The method comprises the following steps: firstly, constructing a generalized Voronoi graph on an occupied grid map obtained by exploration of multiple unmanned vehicles; secondly, quickly calculating the distance between each unmanned station point and each task point by using the topological grid to obtain a distance matrix; on the basis of the distance matrix, a task planning problem of the unmanned vehicle is modeled into a multi-objective optimization problem, the multi-objective optimization problem is divided into an upper-layer task allocation problem and a lower-layer path sequence decision problem, the upper layer adopts a multi-objective particle swarm optimization algorithm, tasks are reasonably allocated to appropriate unmanned vehicles, and the lower layer adopts a multi-objective particle swarm optimization algorithm. For each unmanned vehicle, a trained TSP deep learning solution model is adopted to calculate an optimal execution sequence; on the basis of obtaining a specific scheme, in a heuristic area of a topology network in an environment map, an RRT * algorithm is improved, and a planning specific execution path is obtained. According to the method, rapid path planning is realized, and high-quality cost is provided for task allocation.

Description

Technical Field

[0001] The present invention belongs to the technical field of mission planning for multiple unmanned platforms, and specifically relates to a rapid mission planning technology based on deep learning. Background Art

[0002] With the rapid development of science and technology, unmanned platform technology has made breakthrough progress. Unmanned platforms have shown great application potential in many fields with their excellent adaptability, high intelligence level and good carrying capacity. However, with the deepening of application, the application scenarios have gradually expanded from simple obstacle-free environments to more complex, dark and unknown environments. The complexity of tasks has also increased sharply. Therefore, the collaboration of multiple unmanned platforms has become an inevitable choice for you. The collaborative planning of multiple unmanned platforms also has important practical value. First, the collaborative planning of unmanned systems can improve the efficiency and reliability of task execution. Secondly, in complex environments with unknown and dangerous conditions, the collaboration of multiple unmanned platforms can greatly reduce labor costs and risks. Therefore, whether in military fields such as border defense and combat, or in civilian fields such as agriculture and transportation, the collaborative task planning of unmanned systems has strong research significance.

[0003] Existing research primarily divides task planning into two levels: path planning and task allocation. However, at the path planning level, both search-based and sampling-based algorithms face computational complexity challenges when faced with large and complex environments. While biomimetic algorithms offer strong global search capabilities, they struggle to meet real-time requirements. At the task allocation level, market-based approaches rely heavily on communication when faced with large-scale tasks. Traditional optimization methods, however, have slow iteration speeds and are difficult to apply in dynamic environments. Learning-based methods become overly reliant on training data when the scene changes. Therefore, in today's more complex environments and demanding task contexts, there is an urgent need to integrate and innovate new methods within existing scenarios. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a fast task planning method based on deep learning that realizes fast path planning based on full utilization of grid maps and generalized Voronoi diagrams.

[0005] The purpose of the present invention is to achieve the following technical solution: a fast task planning method based on deep learning, comprising the following steps:

[0006] S1. Environment Construction: The environmental data obtained by multiple unmanned vehicles exploring the unknown environment are integrated and processed to construct an occupancy grid map of the overall environment map.

[0007] S2, environmental processing: constructing a generalized Voronoi diagram based on the occupancy grid map;

[0008] S3. Distance Matrix Extraction: Use the generalized Voronoi diagram to quickly calculate the distance between each unmanned vehicle station and the task point to generate a distance matrix;

[0009] S4. Intelligent Algorithm Solution: The task planning problem is modeled as a multi-objective optimization problem and solved using multi-objective optimization theory. The two objectives are to minimize the total time cost of multiple vehicles and the maximum time cost of a single vehicle.

[0010] The multi-objective optimization problem is divided into the upper-level task allocation and the lower-level path sequence decision problem. The upper-level uses a multi-objective particle swarm optimization algorithm to allocate tasks, while the lower-level uses a pre-trained TSP deep learning solution model to calculate the optimal execution sequence. The multi-objective optimization problem gradually approaches the two objectives during the iterative process, and then finds a solution that meets the requirements from the Pareto solution.

[0011] S5. Path planning: Based on the optimal solution in the Pareto solution set, the improved RRT* algorithm is applied in the heuristic area to generate a collision-free path.

[0012] The specific implementation method of step S2 is as follows: first, an initial distance map is calculated and constructed using the Brushfire algorithm, and then converted into an initialized generalized Voronoi diagram; when the generalized Voronoi diagram is updated, the map is determined to be updated based on the data detected by the sensors on the unmanned vehicle; when the sensors on the unmanned vehicle detect an environmental change, the affected local area is first identified, and then the generalized Voronoi diagram of only that area is updated; if no environmental change is detected, the generalized Voronoi diagram is not updated;

[0013] When constructing a generalized Voronoi diagram, the distance map is first updated using the dynamic Brushfire algorithm. A "lower" wavefront is then initiated for newly appearing obstacles in the environment, expanding outward from the newly added obstacle to update the distance to the nearest obstacle in the surrounding free grids. A "raise" wavefront is initiated for removed obstacle areas in the environment, starting from the removed obstacle, checking the distance to the nearest obstacle in the neighboring grids and updating it. This results in an updated GVD diagram, which is then subjected to a skeleton-refining pruning operation.

[0014] In step S3, a breadth-first search is first used to find the cell closest to the skeleton of the generalized Voronoi diagram from the starting point and the target point respectively, and the corresponding shortest path is obtained; the A* algorithm is used to find the shortest path on the generalized Voronoi diagram and the shortest path is included in the distance matrix.

[0015] The TSP deep learning solution model adopts an encoder-decoder structure.

[0016] The specific method of improving the RRT* algorithm is as follows: based on the original RRT* sampling, the target bias strategy is used to adjust the sampling direction and step size to improve the RRT algorithm; first, in the sampling direction, the segmented target x of the heuristic path is combined stage_goal The attraction of F att The repulsive force F with the surrounding obstacle area shown by the dotted line rep1 and F rep2 Generate new sampling directions.

[0017] The beneficial effects of this invention are: by fully utilizing grid maps and generalized Voronoi diagrams, it enables rapid path planning, providing high-quality results for task allocation. The dual-layer optimization structure employed at the task allocation level leverages the rapid solution capabilities of deep learning, accelerating the iteration process of the multi-objective optimization algorithm and achieving a significant increase in planning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flowchart of the fast task planning method based on deep learning of the present invention;

[0019] Figure 2 Flowchart of an algorithm for efficiently constructing Voronoi diagrams based on incremental updates;

[0020] Figure 3 Schematic diagram of heuristic paths and regions;

[0021] Figure 4 This is the structure diagram of the TSP solver model based on deep learning;

[0022] Figure 5 To improve the coding and initialization diagram of the multi-objective particle swarm algorithm;

[0023] Figure 6 Schematic diagram of RRT* sampling within the heuristic area. DETAILED DESCRIPTION

[0024] The present invention constructs a generalized Voronoi diagram on the occupancy grid map obtained by exploration of multiple unmanned vehicles to obtain a topological grid that runs through the entire environment. The topological grid is used to quickly calculate the distance between each unmanned vehicle station and each task point to obtain a distance matrix. The task planning problem of the unmanned vehicle is then modeled as a multi-objective optimization problem, which is divided into the upper-level task allocation and the lower-level path sequence decision problem. The upper layer uses a multi-objective particle swarm optimization algorithm to reasonably allocate tasks to appropriate unmanned vehicles. For each unmanned vehicle, a trained TSP deep learning solution model is used to calculate the optimal execution order. The algorithm is iterated multiple times to achieve fast and accurate task planning. Finally, based on the specific solution, the RRT* algorithm is improved within the inspiration area of ​​the topological network in the environmental map to achieve a collision-free and reasonable path, and the specific execution path is planned.

[0025] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, a fast task planning method based on deep learning of the present invention includes the following steps:

[0027] S1. Environment Construction: The environmental data obtained by multiple unmanned vehicles exploring the unknown environment are integrated and processed to construct an occupancy grid map of the overall environment map.

[0028] S2. Environment processing: Construct a generalized Voronoi diagram based on the occupancy grid map; In this embodiment, an efficient algorithm based on incremental update is used to dynamically construct the generalized Voronoi diagram (GVD diagram). The core idea is to combine the construction process of the generalized Voronoi diagram with the update process of the occupancy grid map, and avoid global reconstruction through local update strategy, thereby significantly improving the computational efficiency. The specific process is as follows: Figure 2 As shown in the figure, the initial distance map (DM) is first constructed using the Brushfire algorithm and converted into an initialized generalized Voronoi diagram. When updating the generalized Voronoi diagram, the data detected by the unmanned vehicle's sensors determines whether the map needs to be updated. When the unmanned vehicle's sensors detect an environmental change, they first identify the affected local area and then update the generalized Voronoi diagram of only that area, rather than recalculating the Voronoi diagram for the entire map. If no environmental change is detected, the generalized Voronoi diagram is not updated. When constructing the generalized Voronoi diagram, the distance map is first updated using the dynamic Brushfire algorithm. A "lower" wavefront is then initiated for newly appearing obstacles in the environment, expanding outward from the newly added obstacle to update the distance to the nearest obstacle in the surrounding free grids. A "raise" wavefront is initiated for removed obstacle areas in the environment, starting from the removed obstacle and checking that the nearest obstacle distance of neighboring grids is the removed obstacle, and then updating the distance. The updated GVD diagram is then obtained. To eliminate redundant connections and maintain the sparsity of the mesh, the updated GVD graph is pruned with skeleton refinement. This incremental update not only effectively handles temporary obstacles in dynamic environments but also significantly reduces computational overhead.

[0029] S3. Distance matrix extraction: In order to greatly shorten the path cost solution time, the generalized Voronoi diagram is used to quickly calculate the distance between each unmanned vehicle station and the task point to generate a distance matrix. Figure 3As shown in the figure, first, a breadth-first search is used to find the cell closest to the skeleton of the generalized Voronoi diagram from the starting point and the target point (the purple point in the figure, the starting point in the lower left corner and the target point in the upper right corner), and the corresponding shortest path is obtained, which is the blue curve part in the figure; after obtaining the starting point and end point cells on the generalized Voronoi diagram, the A* algorithm is used to find the shortest path on the sparse Voronoi diagram (corresponding to the red line part in the figure), so that a feasible path is quickly obtained and included in the distance matrix.

[0030] S4. Intelligent Algorithm Solution: The task planning problem is modeled as a multi-objective optimization problem and solved using multi-objective optimization theory. The ultimate goal of the algorithm is to rationally assign tasks to multiple autonomous vehicles and find conflict-free matching of tasks to minimize global costs. Therefore, the two objectives are to minimize the total time cost of multiple vehicles and the maximum time cost of a single vehicle.

[0031] The constraints are: (1) each task must be executed and can only be executed once; (2) each robot should start from the corresponding robot station and return to the robot station after completing the task;

[0032] The multi-objective optimization problem is divided into the upper-level task allocation and the lower-level path sequence decision problem; the upper level uses the multi-objective particle swarm optimization algorithm to allocate tasks, and the lower level uses the pre-trained TSP deep learning solution model to calculate the optimal execution order; the multi-objective optimization problem gradually approaches the two objectives in the iterative process, and then finds a solution that meets the requirements from the Pareto solution; multi-objective optimization is an overall mathematical modeling method, and the upper and lower-level tasks are the specific methods of the optimization algorithm, that is, a deep learning model is embedded in the multi-objective optimization algorithm to quickly decide the specific execution order of each unmanned vehicle.

[0033] The path length between the unmanned vehicle base station and each task point is obtained through the path planning algorithm, that is, the adjacency matrix (Adjacent matrix), so that the underlying sequential decision problem is modeled as a traveling salesman problem (TSP). The nodes are the task points, and the edges are the specific paths planned between two points. Drawing on the encoder-decoder structure, a graph neural network combining dynamic attention mechanism and gated residual connection is proposed. Its structure is as follows Figure 4 shown.

[0034] In the encoder part, the node feature v is first i and non-Euclidean distance edge features e ij Perform embedding processing to obtain a high-dimensional embedding representation of the information. The specific formula is expressed as:

[0035]

[0036] Where W v and W e is the learnable weight matrix, b v and b e is the bias term of the two features;

[0037] Then it is sent to the multi-layer dynamic residual attention network for feature learning. The multi-layer dynamic residual attention network includes L layers of stacked dynamic residual attention networks. Each network is represented as:

[0038]

[0039] DRGAT stands for Dynamic Residual Attention Network, Represents the node features of the l-1 layer;

[0040] In each layer of the dynamic residual attention network, the dynamic graph attention layer and the gated recurrent unit are included in sequence. The output and input of the dynamic graph attention layer are residually connected and then input into the gated recurrent unit. Each layer of the dynamic residual attention network calculates the attention coefficient based on the node features and edge features of the previous layer input. The multi-head attention mechanism based on the dynamic graph attention network obtains the updated features of the graph neural network through message passing The gating mechanism is designed to effectively alleviate the gradient vanishing problem in deep networks while retaining historical feature information and balancing the importance of historical features and current features. Its principle is to introduce the new features generated by GRU dynamic fusion attention and the old features of the previous layer in the process of fusing edge feature extraction into node features.

[0041] The decoder uses a Transformer-based attention mechanism to generate path-sequential embeddings from the encoder output. This process involves two steps: global path-sequential planning and local node selection. The decoder consists of a sequential context embedding module, a multi-head attention mechanism, a single-head attention mechanism, and a softmax function. The output of the softmax function is fed back into the input of the context embedding module.

[0042] First, the global path sequence planning is based on the node embedding x of the softmax feedback g and the node features output by the encoder Calculate the initial context vector through the context embedding module The initial context is then further processed through a multi-head attention mechanism Generate new context Among them, MHA represents the multi-head attention mechanism. represents the j-th node feature, Q k Represents the query weight matrix, K kRepresents the weight matrix of the key, V k The weight matrix representing the value (Value); based on the updated context vector c t and node embedding, the compatibility between the current context embedding and the embedding of all unvisited nodes is calculated through the single-head attention mechanism to generate the attention weight; if j has not been visited, it can be calculated based on Calculate the compatibility score of unvisited nodes, where d k is the dimension of the key vector, is a learnable projection matrix used to transform the updated context vector c t Mapped to query vector; is the projection matrix, which is used to map node features into weight vectors; the compatibility score of the visited nodes is set to -∞; finally, the compatibility score is converted into a probability distribution p through softmax jt =softmax(u cj ), the node with the highest probability is selected as the next node. This decoder is pre-trained using reinforcement learning.

[0043] After obtaining the trained TSP solver, the lower-level path sequence decision problem has been solved. In the process of integration with the upper-level algorithm, some adaptive improvements need to be made to the multi-objective particle swarm algorithm. The first step is encoding and particle initialization. A solution to the unmanned vehicle path planning is used as a particle, such as Figure 5 As shown, each row x1,x1,…,x N Each element t1, t1,…, t in each particle represents a particle, N represents the total number of particles; n Each represents a task, and there are n tasks in total; the number 1, 3, 2, ..., 4 corresponding to each task represents the number of the robot that performs the task. The task set R1 of each unmanned vehicle is obtained by real number coding: Logistic mapping is used to generate the initial population. Next, particles are updated and maintained. The NSGA-II crossover and mutation mechanism is introduced to generate new particles based on traditional velocity and position updates, maintaining population diversity. Finally, external archives are updated and maintained. During the fusion process, each iteration of the multi-objective particle swarm algorithm utilizes a deep learning model to calculate the population's fitness and updates the optimal order for each particle. This iterative process yields the Pareto solution of the intelligent algorithm.

[0044] S5. Specific path planning: Based on the optimal solution in the Pareto solution set, the improved RRT* algorithm is applied in the heuristic area to generate a collision-free path. Figure 6As shown, based on the original RRT* sampling, the target bias strategy is used to adjust the sampling direction and step size to improve the RRT algorithm. First, in the sampling direction, the segmented target x of the heuristic path is combined stage_goal The attraction of F att The repulsive force F with the surrounding obstacle area shown by the dotted line rep1 and F rep2 Generate a new sampling direction, such as the resultant force direction F represented by the green line in the figure total , and the sampling step size is dynamically adjusted according to the distance to the nearest obstacle in the map, adapting to the corridor environment of the heuristic area, and finally generating a collision-free path and performing smooth optimization.

[0045] Through the processing of the above steps, the present invention realizes rapid task planning for multiple unmanned vehicles in complex large-scale indoor environments. Compared with traditional task planning methods, the present invention introduces a fast path cost solution technology based on Voronoi and a TSP solution model based on deep learning, which can effectively accelerate the iterative process of the algorithm and show better efficiency and accuracy in more complex scenarios.

[0046] By combining generalized Voronoi diagrams and deep learning with traditional optimization methods at both the path planning and task allocation levels, this invention not only demonstrates high theoretical value but also has numerous practical applications. Experimental verification demonstrated superior solution balance and planning efficiency compared to the classic advanced algorithms NSGA-II and MOACO in scenarios with autonomous vehicles of varying speeds and task completion times.

[0047] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A fast task planning method based on deep learning, characterized in that: The following steps are involved: S1. Environment Construction: The environmental data obtained by multiple unmanned vehicles exploring the unknown environment are integrated and processed to construct an occupancy grid map of the overall environment map. S2, environmental processing: constructing a generalized Voronoi diagram based on the occupancy grid map; S3. Distance Matrix Extraction: Use the generalized Voronoi diagram to quickly calculate the distance between each unmanned vehicle station and the task point to generate a distance matrix; S4. Intelligent algorithm solution: Model the task planning problem as a multi-objective optimization problem and solve it using multi-objective optimization theory; The two objectives are to minimize the total time cost of multiple vehicles and the maximum time cost of a single vehicle; The multi-objective optimization problem is divided into the upper-level task allocation and the lower-level path sequence decision problem. The upper-level uses a multi-objective particle swarm optimization algorithm to allocate tasks, while the lower-level uses a pre-trained TSP deep learning solution model to calculate the optimal execution sequence. The multi-objective optimization problem gradually approaches the two objectives during the iterative process, and then finds a solution that meets the requirements from the Pareto solution. S5. Path planning: Based on the optimal solution in the Pareto solution set, the improved RRT* algorithm is applied in the heuristic area to generate a collision-free path.

2. A fast task planning method based on deep learning according to claim 1, characterized in that: The specific implementation method of step S2 is as follows: first, an initial distance map is calculated and constructed using the Brushfire algorithm, and then converted into an initialized generalized Voronoi diagram; when the generalized Voronoi diagram is updated, the map is determined to be updated based on the data detected by the sensors on the unmanned vehicle; when the sensors on the unmanned vehicle detect an environmental change, the affected local area is first identified, and then the generalized Voronoi diagram of only that area is updated; if no environmental change is detected, the generalized Voronoi diagram is not updated; When constructing a generalized Voronoi diagram, the distance map is first updated using the dynamic Brushfire algorithm. A "lower" wavefront is then initiated for newly appearing obstacles in the environment, expanding outward from the newly added obstacle to update the distance to the nearest obstacle in the surrounding free grids. A "raise" wavefront is initiated for removed obstacle areas in the environment, starting from the removed obstacle and checking the distance to the nearest obstacle in the neighboring grids to match the removed obstacle and updating the distance. This results in an updated GVD diagram, which is then subjected to skeleton refinement and pruning.

3. The method for rapid task planning based on deep learning according to claim 1, wherein: In step S3, a breadth-first search is first used to find the cell closest to the skeleton of the generalized Voronoi diagram from the starting point and the target point respectively, and the corresponding shortest path is obtained; the A* algorithm is used to find the shortest path on the generalized Voronoi diagram and the shortest path is included in the distance matrix.

4. The method for rapid task planning based on deep learning according to claim 1, characterized in that: The TSP deep learning solution model adopts an encoder-decoder structure.

5. The method for rapid task planning based on deep learning according to claim 1, characterized in that: The specific method of improving the RRT* algorithm is as follows: based on the original RRT* sampling, the target bias strategy is used to adjust the sampling direction and step size to improve the RRT algorithm; first, in the sampling direction, the segmented target x of the heuristic path is combined stage_goal The attraction of F att The repulsive force F with the surrounding obstacle area shown by the dotted line rep1 and F rep2 Generate new sampling directions.

Images