Rapid Motor Adaptation

A method that lets quadruped robots sense terrain or load changes and adjust gait within a fraction of a second.

RMA was proposed by Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik at UC Berkeley and Carnegie Mellon University, published at RSS 2021. Real ground varies in softness, friction, load, and motor wear, and a policy trained only in simulation often falls on the real robot. RMA has two parts. A base policy, trained in simulation, has access to these environment parameters (privileged information) and compresses them into a low-dimensional “extrinsics” vector that modulates its actions. An adaptation module sees only a recent window of joint state and action history, and learns to infer that same vector. Deployed on a Unitree A1 with no real-robot fine-tuning, it adapts within a fraction of a second to sand, foam mats, stairs, and slippery ground. This approach — using privileged information during training and estimating it from history at deployment — later became a standard recipe for sim-to-real transfer in legged robots.

ExampleA sudden load is added to a Unitree A1's back; the adaptation module estimates the new extrinsics vector from the recent change in joint states, and the policy immediately adjusts its gait and keeps walking, with no retraining needed.

Also called

RMA

Related

Sim-to-Real Transfer · Privileged Information · Teacher-Student Distillation · RL-based Locomotion Control · Domain Randomization · HORA

Sources

RMA: Rapid Motor Adaptation for Legged Robots 项目主页 (Chinese)

As of

2021-07