Research On Hopping Trajectory Planning For Asteroid Probe Via Deep Reinforcement Learning

The asteroid's micro-gravity environment makes traditional wheeled probes difficult to complete asteroid surface survey tasks,and the use of hopping probes can easily cross various obstacles,can be well adapted to the asteroid environment,and lower Energy consumption probes a larger range.However,there are few researches on the hopping trajectory planning in the related research of hopping asteroid probes,especially the research on the continuous hopping trajectory planning of the probes lacking large-scale transfer.Therefore,after deep research on deep reinforcement learning,this paper applies it to the study of hopping trajectory planning,designs corresponding artificial neural networks for flat ground conditions and obstacle avoidance tasks,uses deep reinforcement learning algorithms for training,and conducts simulation test verification..The main research contents of the paper are as follows:The asteroid probe is designed as an ideal rigid cube with a three-axis orthogonal flywheel mounted at the center of mass.The hopping process of the probe on the surface of the asteroid is modeled and the hopping movement simulation under uncontrolled conditions is completed.Analysis shows that the pre-collision state of the probe is the decisive factor in changing the trajectory.Then the simple hopping trajectory control strategy based on the velocity direction of the collision point is analyzed.It is pointed out that the steering angle is difficult to control precisely,and the speed will be reduced due to the collision during long-distance hopping movement.It is difficult to complete the hopping movement task,so the probe needs further study for hopping trajectory planning.Based on the deep research of the deep reinforcement learning algorithm,this paper designs a corresponding network structure based on the deep deterministic strategy gradient algorithm framework and learns the hopping trajectory planning strategy of the probe.Taking into account the conflict between the large amount of data required for deep reinforcement learning and the high computational time required for simulation calculations,this paper will rationally simplify the air flight process that is not important in the task of trajectory planning for the probe's hopping trajectory,effectively reducing the time required for simulation..If only rewarding the probe to reach the target,the reward value will be too sparse,which is not conducive to training.Therefore,this paper designs a reasonable single-step reward value for the hopping mobile mission and accelerates the training process.This paper first regards the asteroid surface as an ideal flat surface,adopts a feedforward neural network as a strategy network,and uses depth-enhanced learning algorithms to train a large amount of simulation data.It shows good performance in simulation tests and can effectively Any initial position and speed complete the hopping to the target point.Then we consider the rugged terrain on the surface of asteroid.This paper regards it as an obstacle area that the probe should avoid to contact with the ground.It requires the hopping trajectory planning strategy to have obstacle avoidance capability.Based on this,this paper introduces the idea of value iteration network,uses the value iterative network to process the grid map with the location information of the obstacle area,extracts the effective obstacle avoidance features,and fuses with the rest of the probe features.After training the strategy network embedded with value iterative network,the simulation test results show that by introducing a value iteration network,the strategy learned by the network has a certain ability to avoid obstacles,and its performance in all tasks exceeds the aforementioned feedforward neural network.