Stability-Guaranteed Reinforcement Learning for Contact-Rich Manipulation

Reinforcement learning (RL) has had its fair share of success in contact-rich manipulation tasks but it still lags behind benefiting from advances robot control theory such as impedance and stability guarantees. Recently, the concept variable (VIC) was adopted into RL with encouraging results. However, more important issue remains unaddressed. To clarify challenge stable RL, we introduce term all-the-time-stability that unambiguously means every possible rollout should be certified. Our contribution is a model-free method not only adopts VIC also achieves all-the-time-stability. Building on recently proposed controller policy parameterization, novel search algorithm inspired by Cross-Entropy Method inherently guarantees stability. experimental studies confirm feasibility usefulness guarantee features, to best our knowledge, first successful application benchmark problem peg-in-hole.