| Asteroid exploration has gradually become a new-hot-trend in the field of deep space exploration.As people know little about asteroids,it is especially important to carry out all-round detection,including surface detection.Asteroids are much smaller in size and mass than planets.The microgravity gravitational field environment poses a great challenge to the wheel detectors commonly used at this stage.Therefore,it is proposed to use a jump detector to perform detection mission on asteroid surfaces.At present,the research on asteroid landing detectors at home and abroad is still in the initial stage,and there are few researches on the planning strategy of the detector’s continuous multiple jump process.Because deep reinforcement learning has both the ability of deep learning to express things and the ability of learning to solve problem-solving strategies,the neural network that responds to the depth-deterministic strategy gradient algorithm design is used to plan the complete process of asteroid probe jumping.The main contents of the paper are as follows:Firstly,the basic modeling of the hopping asteroid detector is carried out,and a simple fast energy exchange strategy is proposed to verify that the detector can realize energy conversion through the collision process.Then,based on the depth deterministic gradient algorithm,the neural network and the reward function are designed.The collision planning strategy is studied without considering the flywheel control ability,and it has a good performance in the test.At the same time,in order to test the robustness of the collision motion planning strategy of the neural network in the case of small interference ground,it is tested under the two environments of small angle slope and random soil information.Excellent performance.In order to study the climbing ability of the detector motion process,the neural network structure was redesigned in a large angle slope environment,and the slope information was used as a separate input,combined with the detector collision motion planning strategy,to complete the training.The new neural network performs the bevel environment test at different angles,and compares and tests the performance of the original neural network under the same slope environment.The overall trend of performance changes with the increase of the tilt angle,but under the large angle slope environment the performance of the new neural network is significantly better than that of the original neural network.In addition,the neural network is modeled,trained and tested in a spherical ground environment.Even if the representation of the detector state space is changed,it still does not affect the learning of the collision motion planning strategy.Considering the flywheel control ability,the flywheel information is input into the neural network as part of the detector state space,and it is found that it is difficult to learn the balance between the collision motion planning strategy and the flywheel control capability.Therefore,a flywheel unloading is proposed.The planning strategy is to unload the flywheel during the collision of the detector with the ground.Through the redesign of the state space and the reward function,the neural network can better learn the flywheel unloading planning strategy and also perform well in the range of motion indicators. |
