Deep Reinforcement Learning Based Trajectory Planning for Hopping on Low-Gravity Asteroid Surface

In a small-body exploration mission, the rover may deviate from the expected target point due to the dispersion of delivery. This paper proposes a hopping trajectory planning method based on deep reinforcement learning to achieve a precision landing on a low-gravity surface. First, the dynamic model of the hopping rover is established. Then, the hopping scheme is proposed with the attitude angle and angular velocity as control variables. In order to rapidly solve the control variables, the deep reinforcement learning algorithm is utilized for the autonomous hopping trajectory planning. The landing process is divided into an approach and a deceleration stage, and two agents are trained according to the reward functions of the two stages. To achieve the expected attitude angle and angular velocity given by the agents' outputs, the control torque is solved using sliding mode control method. Finally, the hopping trajectory planning method are verified in a landing mission on low-gravity surface. The results show that the rover can reach and stop at the target by intelligent hopping under various initial conditions.