Residual Reinforcement Learning for Robotic Assembly of Large-Scale Aerospace Components

Abstract

Robotic assembly of large-scale aerospace components demands millimeter-level accuracy under intermittent contacts, while collecting rich interaction data remains costly and risky. This paper presents a demonstration-guided residual reinforcement learning framework for precision assembly. A diffusion-based action-chunking policy trained from limited teleoperated demonstrations generates long-horizon nominal trajectories at a low frequency. A closed-loop residual policy optimized with PPO then adds per-step pose corrections to compensate for distribution shift and contact dynamics during the final mating phase. An action-hold sparse reward is introduced to promote stable mating rather than transient contact. Simulation experiments on KUKA KR210 industrial robot demonstrate that the proposed approach improves assembly success rate and efficiency compared with baselines, validating the effectiveness of combining imitation-based priors with closed-loop residual refinement.