Reinforcement Learning of Phase-Aware Admittance Control Parameters for Robotic Peg-in-Hole Assembly

Robotic peg-in-hole assembly remains a challenging problem in automation, primarily due to uncertainties caused by positioning errors and complex contact dynamics. Conventional admittance control, which lacks parameter self-adaptation capabilities, often fails to meet the varying dynamic requirements across different assembly phases. To address this issue, this paper proposes a phase-aware admittance control framework based on reinforcement learning. The method adopts a hierarchical architecture: a low-level parameterized admittance controller governs physical interactions, while a high-level agent, based on the Soft Actor-Critic (SAC) algorithm, dynamically tunes the admittance parameters to enable policy adaptation. The framework is validated in CoppeliaSim, where a dense reward function incorporating guidance and progress metrics is designed to accelerate training. Experimental results show that the proposed approach significantly outperforms fixed-parameter control and other baselines, achieving a success rate exceeding 98% along with superior stability and convergence. This study demonstrates the feasibility of optimizing classical control parameters via reinforcement learning, providing an efficient and stable solution for complex contact-intensive tasks.