"Rapid Motor Adaptation" Algorithm Lets Legged Robots Quickly Adapt to Difficult Terrain
Trained entirely in simulation and deployed without fine-tuning, RMA proves state-of-the-art — even for terrain the robot's never seen.
A quartet of roboticists from the UC Berkeley, Carnegie Mellon University, and Facebook's AI Research division have unveiled a new algorithm designed to help legged robots adapt to uneven and changing terrains: Rapid Motor Adaption, or RMA.
"Successful real-world deployment of legged robots would require them to adapt in real-time to unseen scenarios like changing terrains, changing payloads, wear and tear," the team writes in the abstract to the paper. "This paper presents Rapid Motor Adaptation (RMA) algorithm to solve this problem of real-time online adaptation in quadruped robots."
"RMA consists of two components: a base policy and an adaptation module. The combination of these components enables the robot to adapt to novel situations in fractions of a second. RMA is trained completely in simulation without using any domain knowledge like reference trajectories or predefined foot trajectory generators and is deployed on the A1 robot without any fine-tuning."
The paper, brought to our attention by TechCrunch, shows impressive performance from the algorithm — double impressive, given its training takes place entirely in simulation and lack of environment-specific tuning for deployment. The RMA-equipped robot is shown to be able to traverse a range of surfaces without difficulty — including a 90% success rate in walking across a plastic surface when the researchers had poured oil over its feet.
Its success at walking down stairs was somewhat lower, at 70% — but the researchers are still declaring it a success, "given that the robot never sees a staircase during training."
Full details on RMA are available on the website and in the linked paper.
