Reproduce “Learning Arm-Assisted Fall Damage Reduction and Recovery for Legged Mobile Manipulators” on OpenResearch

About this paper

Yuntao Ma, Farbod Farshidian, Marco Hutter

A learning-based framework from ETH Zürich enables legged mobile manipulators to reduce fall damage and recover using their robotic arm, employing an asymmetric actor-critic reinforcement learning setup. The system achieves lower base contact impulse and acceleration during falls, demonstrated by a 98.9% success rate for arm-assisted recovery on hardware.

Key insights

• The asymmetric actor-critic architecture, utilizing privileged observations for the critic, is essential for accurately estimating the value function in finite-horizon tasks with time-varying rewards, even with a minimally-observed actor.
• A time-invariant actor policy, which does not observe the remaining episode time, leads to more robust and temporally consistent recovery behaviors compared to a time-variant actor that might exploit reward timing for sub-optimal, abrupt actions.
• Leveraging the robotic arm for fall mitigation and recovery not only significantly increases recovery success rates but also reduces the torque burden on the robot's legs, making the recovery process more efficient.

Summary by alphaXiv.