2020/04/22 by Shahbaz Abdul Khader, Khader, Shahbaz A., Hang Yin +5
Engineering · #FOS: Computer and information sciences #Iterative Learning Control Systems #Muscle activation and electromyography studies #Robot Manipulation and Learning #Robotics (cs.RO)
paper · pdf · doi:10.48550/arxiv.2004.10886
openalex publication_date 2020/04/22 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Reinforcement learning (RL) has had its fair share of success in contact-rich\nmanipulation tasks but it still lags behind in benefiting from advances in\nrobot control theory such as impedance control and stability guarantees.\nRecently, the concept of variable impedance control (VIC) was adopted into RL\nwith encouraging results. However, the more important issue of stability\nremains unaddressed. To clarify the challenge in stable RL, we introduce the\nterm all-the-time-stability that unambiguously means that every possible\nrollout will be stability certified. Our contribution is a model-free RL method\nthat not only adopts VIC but also achieves all-the-time-stability. Building on\na recently proposed stable VIC controller as the policy parameterization, we\nintroduce a novel policy search algorithm that is inspired by Cross-Entropy\nMethod and inherently guarantees stability. Our experimental studies confirm\nthe feasibility and usefulness of stability guarantee and also features, to the\nbest of our knowledge, the first successful application of RL with\nall-the-time-stability on the benchmark problem of peg-in-hole.\n