SHaRe-RL: Structured, Interactive Reinforcement Learning for Contact-Rich Industrial Assembly Tasks

Abstract

High-mix low-volume (HMLV) industrial assembly, common in small and medium-sized enterprises (SMEs), requires the same precision, safety, and reliability as high-volume automation while remaining flexible to product variation and environmental uncertainty. Current robotic systems struggle to meet these demands. Manual programming is brittle and costly to adapt, while learning-based methods suffer from poor sample efficiency and unsafe exploration in contact-rich tasks. To address this, we present SHaRe-RL, a reinforcement learning framework that leverages multiple sources of prior knowledge. By (i) structuring skills into manipulation primitives, (ii) incorporating human demonstrations and online corrections, and (iii) bounding interaction forces with per-axis compliance, SHaRe-RL enables efficient and safe online learning for long-horizon, contact-rich industrial assembly tasks. Experiments on the insertion of industrial Harting connector modules with 0.2-0.4 mm clearance demonstrate that SHaRe-RL achieves reliable performance within practical time budgets. Our results show that process expertise, without requiring robotics or RL knowledge, can meaningfully contribute to learning, enabling safer, more robust, and more economically viable deployment of RL for industrial assembly.

Figures (8)

• Figure 1: Structure, interaction and safety. SHaRe-RL is a hybrid approach that combines a hand-designed primitive sequence with interactive learning in selected primitives and adaptive per-axis force limits for bounded contact forces.
• Figure 2: Adaptive force limits under ideal conditions Left: Time response of a stable adaptive limit. Right: Phase-plane cobweb plot illustrating the recurrence in Eq.\ref{['eq:recurrence']}.
• Figure 3: Experimental setup and task. Top: UR3e robot and Harting HanDD industrial connector used for evaluation. Bottom: Cropped policy input over time.
• Figure 4: MP-Net for the HanDD connector insertion. Light-blue and dark-blue arrows indicate the train and deployment loop respectively. The learned termination condition $\lambda(\mathbf{o}_t)$ and the randomized reset MP6 close the loop for repeated trials. The table lists the per-axis controller setpoints for each primitive.
• Figure 5: Main comparison of learning approaches. Success rate, cycle time, and intervention rate vs wall-clock time (mean ± s.e.m., 2 seeds).