Reproduce “Learning Variable Impedance Control for Contact Sensitive Tasks” on OpenResearch

About this paper

Miroslav Bogdanovic, Majid Khadiv, Ludovic Righetti

At the Max-Planck Institute for Intelligent Systems, researchers demonstrated that training reinforcement learning policies with variable joint impedance control and an added trajectory tracking reward significantly improves robustness to contact uncertainties. This approach enabled successful and stable sim-to-real transfer for contact-sensitive tasks, outperforming direct torque and fixed-gain control.

Key insights

• Joint-space variable impedance control serves as a superior action space for reinforcement learning in contact-rich tasks, offering greater robustness than direct torque or fixed-gain control.
• A simple trajectory tracking reward term significantly enhances policy interpretability and enables stable sim-to-real transfer by ensuring learned desired trajectories are physically achievable.
• Optimal impedance modulation for contact tasks involves dynamically reducing gains before impact for absorption and increasing them during force exertion for efficient push-off.

Summary by alphaXiv.