A flying vehicle path planning method based on modal prediction
By adopting modal prediction methods and game learning frameworks in flight vehicle path planning, the problems of air-ground mode switching and multi-modal path planning of flight vehicles in complex mutation environments are solved, and the effective obstacle avoidance and task completion of flight vehicles in the mutation environment are achieved.
Patent Information
- Application Number
- CN202210224659.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-09
- Publication Date
- 2025-04-15
- Estimated Expiration
- 2042-03-09
AI Technical Summary
The prior art lacks planning strategies for air-ground mode switching in complex and mutation environments, especially when the map prior information fails, it is difficult to effectively predict and plan multimodal obstacle avoidance paths.
A method of flight vehicle path planning based on modal prediction is proposed, using a static environment global multimodal path planning and a hierarchical architecture of local multimodal path planning in complex mutation environments, using the game learning framework to collect modal switching training data in a static environment, and using a modal prediction network to predict switching modes in a mutation environment, and finally generating a local multimodal path.
It realizes effective air-ground mode switching decisions and multi-modal path planning for flying vehicles in complex mutation environments, and improves the obstacle avoidance ability and task completion efficiency of flying vehicles in the mutation environment.
Smart Images
Figure CN114545981B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of aircraft, and in particular relates to a flying vehicle path planning method based on modal prediction. Background Art
[0002] Flying vehicles have two motion modes, ground driving mode and flight mode. By switching between the two modes, they can both drive on the ground and fly in the air. When faced with obstacles that are difficult for ground vehicles to climb over, flying vehicles can switch to flight mode to climb over obstacles. And compared to aircraft, flying vehicles do not need to maintain a long-term flight state, they only need to take off when necessary, which effectively improves energy utilization.
[0003] Most existing path planning solutions are for single motion modes of unmanned ground vehicles or unmanned aerial vehicles. There is a lack of a planning strategy for flying vehicles that considers air-to-ground mode switching. Furthermore, flying vehicle path planning strategies that consider modal predictions of sudden environmental changes are more challenging and less studied.
[0004] Modal prediction planning in complex mutation environments has great research value. For example, after an earthquake, houses collapsed and existing map information became completely invalid. This complex mutation environment requires flying vehicles to be able to predict the air-ground mode to be switched based on the perceived local obstacle information, autonomously plan multi-modal obstacle avoidance paths, and complete a series of tasks such as rescue.
[0005] At present, the research on path planning solutions mainly focuses on single-mode research such as unmanned vehicles and unmanned aerial vehicles, and lacks planning strategies for flying vehicles that consider air-ground mode switching.
[0006] Especially in complex mutation environments, the map prior information has become invalid, and the switching decision between air-ground modes lacks the guidance of complete obstacle information. It is necessary to predict the switching action in the current complex mutation environment based on obstacle data training results in a static environment. However, the flying vehicle path planning strategy based on modal prediction is more challenging, and there is less related research.
[0007] Moreover, at present, there is no ready-made static environment training data for flying vehicles, and a plan for collecting relevant data needs to be designed. Summary of the invention
[0008] The present invention proposes a flying vehicle path planning method based on modal prediction, including a global-local hybrid path planning with a hierarchical architecture of global multimodal path planning in static environments and local multimodal path planning in sudden change environments. The method focuses on solving the following three problems: static environment modal switching training data collection and processing, complex sudden change environment air-ground modal switching prediction and flying vehicle multimodal path generation. The specific method is as follows:
[0009] The flying vehicle path planning method based on modal prediction includes a two-layer architecture, the first layer is a global path planning layer for static environments, and the second layer is a local planning layer for complex and sudden change environments;
[0010] The first layer is mainly used to solve the problem of collecting and processing modal training data in a static environment. In the first layer, the vehicle adopts a game learning framework to complete global planning and collects modal switching training data, which includes the relative position information between the vehicle and the obstacle, the obstacle attribute information, and the switching action of the air-ground mode. The relative position information and the obstacle attribute information are the input of the subsequent modal prediction network, and the switching action of the air-ground mode is the output of the modal prediction network.
[0011] The second layer solves the problem of air-to-ground mode switching prediction and multi-modal path planning for flying vehicles in complex mutation environments. The mode prediction network is trained with the training data collected by the first layer. In the mutation environment, the switching mode is predicted based on the relative position information and obstacle attribute information collected by the sensor. Finally, the predicted mode value is sent back to the learning framework of the first layer to obtain the local planning path.
[0012] The first level includes the following four steps:
[0013] Step 1: Obtain the map information of the static environment and the absolute position information of the vehicle for subsequent global path planning.
[0014] Step 2: Taking into full consideration the multi-domain motion capability of the flying vehicle, set the corresponding reward function value for the air-ground mode switching action and the movement in different directions after the switch. With the air-ground mode switching to leap over obstacles, the flying vehicle's passage area is greatly widened. Accordingly, the movement in the direction of the leapable area will not be punished by obstacles. Based on the set reward function, the pursuit game is used to obtain the air-ground mode switching decision at the appropriate time and position. In this game, the flying vehicle's pursuit target vehicle is constructed. The flying vehicle and the target vehicle have different game requirements for the interval distance. The flying vehicle hopes to shorten the interval distance between the two vehicles as much as possible to chase the target vehicle, while the target vehicle hopes to maximize the interval distance to escape the pursuit. Facing obstacles, both vehicles have different mode switching action options. The selected switching action should meet their respective game requirements to the greatest extent. This is the mode switching Nash equilibrium solution. Finally, through the solution of the equilibrium solution of the behavioral game of the two vehicles, the flying vehicle can obtain the appropriate air-ground mode switching decision.
[0015] Step 3: Use the obtained mode switching action for multimodal path planning. During the vehicle's exploration and learning of the environment, game learning pursues the maximum learning reward, and then obtains a global multimodal path that meets the planning goals.
[0016] Step 4: During the planning process, the vehicle records the relative position information with the obstacle, the obstacle attribute information, and the corresponding mode switching information, which are stored for subsequent training.
[0017] The second level includes the following four steps:
[0018] Step 5: In a complex mutation environment, the sensor obtains local perception information of the obstacle, including the relative position information between the vehicle and the obstacle and the obstacle attribute information, as the input of the trained modal prediction network.
[0019] Step 6: The modal prediction network uses the training data collected in the static environment to train the network. After completion, the switching mode is predicted based on the local perception information of obstacles in the complex mutation environment. The predicted modal action is used for the subsequent local path planning.
[0020] Step 7: The predicted modal action is sent back to the first-layer learning framework. The switching action of the current obstacle is known, and we only need to find the motion trajectory that maximizes the learning reward. Finally, we can plan the local multimodal path of the flying vehicle in a complex mutation environment.
[0021] The present invention has the following technical effects:
[0022] (1) A flying vehicle path planning method based on modal prediction is proposed, which solves the difficult problems of air-ground modal switching decision and multi-modal path planning in complex and sudden change environments of flying vehicles.
[0023] (2) The game learning framework is used to plan a global multimodal path for the flying vehicle in a static environment, and to solve the problem of collecting and processing training data for the modal prediction network in a complex and sudden environment.
[0024] (3) A comprehensive path planning solution is designed for flying vehicles, from training data collection to modal prediction training to the final generation of multi-modal paths in complex mutation environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1a The flying vehicle of the embodiment is configured in an airborne flight state;
[0026] Figure 1b The flying vehicle of the embodiment is configured in a ground driving state;
[0027] Figure 2 It is a schematic diagram of the game learning process of the present invention;
[0028] Figure 3 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION
[0029] The specific technical solution of the present invention is explained in conjunction with embodiments.
[0030] Figure 1a and Figure 1b It is a flying vehicle configuration, which has two modes: flying in the air and driving on the ground. It is characterized by including a flight control system, a chassis system and a power system. The flight control system adopts an eight-rotor coaxial double-propeller design. Each propeller is equipped with a separate rotor motor. By adjusting the speed of each rotor motor, the attitude change and motion control of the flying vehicle with high degree of freedom during flight are performed. The chassis system adopts a distributed four-wheel independent drive system. The hub motor integrates the drive components, transmission components and braking devices into the wheel hub. The vehicle is directly driven by the four hub motor assemblies integrated into the wheel hub. The power system consists of an engine motor group, a battery pack, an electric regulator, etc., and adopts a series hybrid power form. The rotor motor and the hub motor are supplied with energy by the same power system.
[0031] The flying vehicle path planning method based on modal prediction includes a two-layer architecture, the first layer is a global path planning layer for static environments, and the second layer is a local planning layer for complex and sudden change environments;
[0032] The first layer is mainly used to solve the problem of collecting and processing modal training data in a static environment. In the first layer, the vehicle adopts a game learning framework to complete global planning and collects modal switching training data, which includes the relative position information between the vehicle and the obstacle, obstacle attribute information (height and width), and the switching action of the air-ground mode. The relative position information and obstacle attribute information are the input of the subsequent modal prediction network, and the switching action of the air-ground mode is the output of the modal prediction network. The game learning framework in the first layer is as follows: Figure 2 shown.
[0033] The game learning framework for static environments includes the following four steps:
[0034] Step 1: Obtain the map information of the static environment and the absolute position information of the vehicle for subsequent global path planning.
[0035] Step 2: Taking into full consideration the multi-domain motion capability of the flying vehicle, set the corresponding reward function value for the air-ground mode switching action and the movement in different directions after the switch. With the air-ground mode switching to jump over obstacles, the passable area of the flying vehicle is greatly widened. Accordingly, the movement in the direction of the leapable area will not be punished by obstacles. Based on the set reward function, the pursuit game is used to obtain the air-ground mode switching decision at the appropriate time and position. In this game, the flying vehicle is constructed to chase the target vehicle. The flying vehicle and the target vehicle have different game requirements for the interval distance. The flying vehicle hopes to shorten the interval distance between the two vehicles as much as possible to chase the target vehicle, while the target vehicle hopes to maximize the interval distance to escape the pursuit. Facing obstacles, both vehicles have different mode switching action options. The selected switching action should meet their respective game requirements to the greatest extent. This is the mode switching Nash equilibrium solution. Finally, through the solution of the equilibrium solution of the behavioral game of the two vehicles, the flying vehicle can obtain the appropriate air-ground mode switching decision.
[0036] Step 3: Use the obtained mode switching action for multimodal path planning. During the vehicle's exploration and learning of the environment, game learning pursues the maximum learning reward, and then obtains a global multimodal path that meets the planning goals.
[0037] Step 4: The game learning framework not only plays a role in global planning, but also provides training data for subsequent modal prediction. During the planning process, the vehicle records the relative position information with the obstacle, the obstacle attribute information, and the corresponding modal switching information, which are stored for subsequent training.
[0038] The second layer solves the problem of air-to-ground mode switching prediction and multi-modal path planning of flying vehicles in complex mutation environments. The mode prediction network is trained with the training data collected by the first layer. In the mutation environment, the switching mode is predicted based on the local perception information of obstacles (relative position information, obstacle attribute information) collected by the sensor, and the predicted mode value is finally sent back to the learning framework of the first layer to obtain the local planning path.
[0039] In summary, the overall process of the flying vehicle path planning method based on modal prediction is as follows: Figure 3 shown.
[0040] The global path acquisition process of the static environment marked by the solid line in the figure is the first-level architecture, and the local path acquisition process of the complex mutation environment marked by the dotted line is the second-level architecture. veh is the relative position information between the vehicle and the obstacle, H obs is the height information of the obstacle, w obs is the width information of the obstacle, M f is the static environment space-ground mode switching action obtained by the game learning framework, M pis the air-ground modal prediction action of the sudden change environment obtained by the modal prediction network, s is the current position state of the flying vehicle, s′ is the next position state, a is the action strategy taken by the flying vehicle, and r is the reward value of the action strategy feedback.
[0041] The flying vehicle path planning method based on modal prediction includes the following steps.
[0042] First is the first-level architecture workflow:
[0043] Step 1: Obtain the absolute position information and map information of the vehicle in the static environment. Same as Step 1 of the above game learning framework.
[0044] Step 2: Set the reward function for different modes of action and movement direction of the flying vehicle, and use the pursuit game to obtain the air-ground mode switching decision at the right time and position. The details are the same as the above game learning framework Step 2.
[0045] Step 3: Plan a multimodal global path in a static environment. The details are the same as Step 3 of the above game learning framework.
[0046] Step 4: Obstacles and mode switching action data in the global planning process of the static environment are collected and stored in the data storage unit for subsequent mode prediction network training. The details are the same as Step 4 of the game learning framework.
[0047] The following is the second-level architecture workflow:
[0048] Step 5: In a complex mutation environment, the sensor obtains local perception information of the obstacle, including the relative position information between the vehicle and the obstacle and the obstacle attribute information, as the input of the trained modal prediction network.
[0049] Step 6: The modal prediction network uses the training data collected in the static environment to train the network. After completion, the switching mode is predicted based on the local perception information of obstacles in the complex mutation environment. The predicted modal action is used for the subsequent local path planning.
[0050] Step 7: The predicted modal action is sent back to the first-layer learning framework. The switching action of the current obstacle is known, and we only need to find the motion trajectory that maximizes the learning reward. Finally, we can plan the local multimodal path of the flying vehicle in a complex mutation environment.
Claims
1. A flying vehicle path planning method based on modal prediction, characterized in that: It includes two layers of architecture: the first layer is the global path planning layer for static environments, and the second layer is the local planning layer for complex and sudden change environments. The first layer is used to solve the problem of collecting and processing modal training data in a static environment. In the first layer, the vehicle adopts a game learning framework to complete global planning and collects modal switching training data, which includes the relative position information between the vehicle and the obstacle, the obstacle attribute information, and the switching action of the air-ground mode. The relative position information and the obstacle attribute information are the input of the subsequent modal prediction network, and the switching action of the air-ground mode is the output of the modal prediction network. The first level includes the following four steps: Step 1: Obtain the map information of the static environment and the absolute position information of the vehicle for subsequent global path planning; Step 2: Taking into full consideration the multi-domain motion capability of the flying vehicle, set the corresponding reward function value for the air-ground mode switching action and the movement in different directions after the switch; by virtue of the air-ground mode switching to jump over obstacles, the flying vehicle's passage area is greatly widened, and accordingly, the movement in the direction of the jumpable area will not be punished by obstacles; based on the set reward function, the pursuit game is used to obtain the air-ground mode switching decision at the appropriate time and position; in this game, the flying vehicle's pursuit target vehicle is constructed, and the flying vehicle and the target vehicle have different game requirements for the interval distance. The flying vehicle hopes to shorten the distance between the two vehicles as much as possible to chase the target vehicle, while the target vehicle hopes to maximize the distance to escape the pursuit; When facing obstacles, both vehicles have different options for mode switching actions. The selected switching actions should meet their respective game needs to the greatest extent. This is the mode switching Nash equilibrium solution. Finally, through the behavioral game of the two vehicles to solve the equilibrium solution, the flying vehicle can obtain the appropriate air-ground mode switching decision. Step 3: Use the obtained mode switching action for multimodal path planning. In the process of the vehicle's exploration and learning of the environment, game learning pursues the maximum learning reward, and then obtains a global multimodal path that meets the planning goal; Step 4: During the planning process, the vehicle records the relative position information with the obstacle, the obstacle attribute information, and the corresponding mode switching information, which are stored for subsequent training; The second layer solves the problem of air-to-ground mode switching prediction and multi-modal path planning for flying vehicles in complex mutation environments. The training data collected by the first layer is used to train the mode prediction network. In the mutation environment, the switching mode is predicted based on the relative position information and obstacle attribute information collected by the sensor. Finally, the predicted mode value is sent back to the learning framework of the first layer to obtain the local planning path. The second layer includes the following three steps: Step 5: In a complex mutation environment, the sensor obtains local perception information of obstacles, including the relative position information between the vehicle and the obstacle and the obstacle attribute information, which serves as the input of the trained modal prediction network; Step 6: The modal prediction network uses the training data collected in the static environment to perform network training. After completion, the switching mode is predicted based on the local perception information of obstacles in the complex mutation environment; the predicted modal action is used for the subsequent local path planning; Step 7: The predicted modal action is sent back to the first-layer learning framework. The switching action of the current obstacle is known, and we only need to find the motion trajectory that maximizes the learning reward. Finally, we can plan the local multimodal path of the flying vehicle in a complex mutation environment.