Deep Reinforcement Learning for Smooth Guidance of a Free-Floating Robot in Point-to-Point Tasks

This work presents a Deep Reinforcement Learning (DRL)-based controller for coordinated control of a floatingbase robotic manipulator system in point-to-point tasks. The system considers a 7-degree-of-freedom manipulator that is mounted onto a free-floating platform which is actuated in rotation via three orthogonal reaction wheels. This paper extends previous literature to full end-effector pose-control tasks. The DRL agent is trained to complete maneuvers from arbitrary initial conditions in terms of both target position and attitude. The DRL agent directly maps the Cartesian task space to the robot’s joint space, without relying on offline path planning or inverse kinematics. The DRL controller commands low-level controllers in a feedback-control setting. It is demonstrated that the frequency at which the DRL agent runs significantly affects the reaching accuracy. The performance of the controller is established via Monte-Carlo simulations. Training the DRL controller at a lower frequency is found to improve performance in this task. It is shown that the DRL policy can be induced to provide smoother control signals through deployment to a higher control frequency after training.