An End-to-End Trajectory Planning Strategy for Free-floating Space Robots

The traditional trajectory planning methods of free-floating space robots have the dynamic singular problem and the difficulty of accurate modeling. Although learning-based approaches have achieved remarkable performance on such task, they mainly focus on single modular design such as the perception, planning, or control part. Optimization-based end-to-end method can well combine perception, planning and control, which not rely on parameters of the dynamic model and reduce the difficulty of manually adjusting modular controllers’ parameters. Therefore, we developed an end-to-end trajectory planning strategy based on optimization with multiple constraints. The whole strategy consists of several multi-layer neural networks and is optimized by a deep reinforcement learning algorithm based on maximum entropy. The results of visualization show that our strategy can capture the information of robotic arm from vision directly. Moreover, we evaluate the kinematic and dynamic features of the system and testify our strategy in anti-interference experiments. The performance of our strategy demonstrates the availability and robustness of the system.