This project investigates how model-based reinforcement learning can be applied to acquire manipulation skills for performing assembly tasks in different application domains of space robotics. The developed approach aims to facilitate self-supervised robot learning that is deployable to various real robot systems based on experience collected inside realistic simulation environments. Skill transfer among application domains will be accomplished by employing high-dimensional actions and observations that are invariant to the environment, task, robot kinematics and visual sensors. The applicability of the proposed approach will be evaluated on real robots under laboratory conditions of planetary and orbital scenarios, demonstrating the feasibility of sim-to-real transfer.
Duration: October 1, 2021 – September 31, 2024 (3 years)