Abstract: Robot learning algorithms have promoted the development of motion planning and control, and a key issue in robot learning is how to building a high performance robot physics engine. Due to the special environment of spacecrafts, characterized by limited sample data and costly experimental conditions, this paper presents a high fidelity simulation environment for spacecrafts with robot learning algorithms. Adhering to the standard Gym framework, the simulation environment supports a variety of mainstream robot learning algorithm libraries and Gym style control/learning algorithms. Utilizing experimental data from the microgravity simulation system, a data driven approach is employed to construct a spacecraft dynamics model for state updates within the simulation environment. As an illustrative example, the mainstream reinforcement learning algorithm Soft Actor Critic is trained and tested in the constructed simulation environment for the spacecraft stabilization task, demonstrating the feasibility of the simulation environment for robot learning algorithm.

Key words: space robot, robot learning, micro low gravity environment simulation, hardware in the loop simulation