Reproduce “A Contact-Safe Reinforcement Learning Framework for Contact-Rich Robot Manipulation” on OpenResearch

About this paper

Xiang Zhu, Shucheng Kang, Jianyu Chen

Researchers from Tsinghua University developed a contact-safe reinforcement learning framework for contact-rich robot manipulation, integrating a high-rate controller with a low-rate RL policy to ensure low contact forces on both the end-effector and the robot arm during interaction and unexpected collisions. The framework achieved end-effector contact forces below 5N and reduced unexpected arm collision forces by approximately three times compared to a baseline in real-world robot experiments.

Key insights

• A combined hierarchical control strategy allows for both high-level task learning via RL and low-level, real-time safety enforcement against diverse contacts.
• Estimating external torques on specific robot links using a momentum observer, distinct from end-effector forces, enables targeted compliant responses to unexpected arm collisions.
• Advanced null-space projection techniques can effectively integrate compliant posture control with primary manipulation tasks, ensuring safety without compromising task performance.

Summary by alphaXiv.