An ensemble reinforcement learning framework for robotic high-precision peg-in-hole assembly via human demonstrations

AbstractFor robotic assembly automation in intelligent manufacturing, high-precision peg-in-hole assembly constitutes a fundamental yet challenging task, where tiny pose errors can cause large contact forces and jamming. This paper proposes an ensemble robotic assembly skill-learning framework that integrates imitation learning (IL) and reinforcement learning (RL), with a geometric-constraint-based automatic pose refinement method to adapt a small number of human demonstrations to robot proprioceptive skill reproductions. By defining insertion-related variables, we train an offline multilayer perceptron (MLP) network to obtain the initial IL policy and then construct an ensemble RL architecture with an action layer generating soft actor–critic (SAC) exploratory actions and a parameter layer outputting hybrid force-position control coefficients and fusion weights that combine IL, hybrid force-position and SAC actions into task-execution commands. We first learn in simulation and then transfer to the real robot, where we first update the IL network with frozen RL parameters and subsequently fine-tune the RL policy with a few real-world episodes. Experiments demonstrate rapid sim-to-real convergence on round-hole assembly and robust cross-geometry transfer from round to triangular and square holes, as well as from square holes to USB insertion. Ablation studies and comparisons with state-of-the-art methods confirm improved sample efficiency, lower insertion forces, and higher success rates for cross-geometry generalization.