Reproduce “Discovering Self-Protective Falling Policy for Humanoid Robot via Deep Reinforcement Learning” on OpenResearch

About this paper

Diyuan Shi, Shangke Lyu, Donglin Wang

The authors train a falling policy for a 29-degree-of-freedom Unitree G1 humanoid in simulation using pushes, fall directions, and actuator failures, so the robot learns to brace rather than hold a prescribed pose. The policy forms an arm triangle and spreads legs. In simulation, it outperforms the compared policies on motion energy and contact force, but not actuation impulse; hardware evidence is qualitative.

Key insights

• Deep Reinforcement Learning enables humanoid robots to self-discover highly effective, robot-specific falling policies, such as a “triangle” bracing structure, that are superior to human-engineered or human-prior-based methods for rigid bodies.
• General-purpose locomotion policies can exacerbate damage during falls, highlighting the critical need for dedicated, actively learned fall protection strategies.
• Strategic damage distribution is key; the learned policy reduces impact on vital body parts by actively engaging and directing forces to more robust, 'sacrificial' components like wrists and shoulders.

Summary by alphaXiv.