Learning-Based Task Space Trajectory Planning Frame- Work With Preplanning and Postprocessing for Uncertain Free-Floating Space Robots

Abstract

This article addresses the challenge of the task space trajectory planning problem for free-floating space robots (FFSRs) with model uncertainties. To ensure the end-effector of the uncertain robot follows a desired trajectory in the task space, a composite planning framework combining preplanning and postprocessing is proposed. The adaptive pseudospectral method-based preplanning exploits the nominal part of the uncertain robot, and considers the dynamics coupling of the nominal system to generate baseline trajectories. These baseline trajectories serve as references for the postprocessing. The reinforcement learning-based postprocessing introduces random system parameters into the training process to improve planning accuracy under model uncertainties. Numerical simulations and experiments conducted on an air-bearing testbed verify the effectiveness of the proposed planning framework for uncertain FFSRs.