In context
Robotic peg-in-hole assembly remains a core challenge in manufacturing automation, especially in unstructured environments where fixed-parameter controllers fail. By 2023, reinforcement learning (RL) was emerging as a promising approach to enable robots to learn assembly skills autonomously, but safety and efficiency during training were still open issues.
What was reported
Researchers from Guangxi University proposed a deep deterministic policy gradient (DDPG)-based variable admittance control algorithm with a fuzzy reward mechanism for robotic peg-in-hole assembly. The method establishes a mechanical model of contact states to guide strategy formulation, then uses an admittance controller for compliant assembly. DDPG online identifies optimal controller parameters, while fuzzy rules in the reward function prevent local optima and improve assembly quality.
Experiments were conducted on holes of five different diameters. The proposed algorithm outperformed fixed-parameter admittance models, completing assembly within 10 steps after convergence. The approach does not require precise modeling of physical contact or prior human knowledge, making it suitable for autonomous operation in unstructured environments.
Why it mattered
This work demonstrated that combining RL with variable admittance control can enhance adaptability and safety in precision assembly tasks, offering a pathway toward more flexible automation in manufacturing where part variability and environmental uncertainty are common.
The proposed algorithm is expected to meet the requirements of autonomous manipulation in unstructured environment.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2023-05-09 · “基于强化学习的机器人轴孔装配算法”