Walk These Ways

A quadruped locomotion controller that learns many gaits in one policy and can be retuned live at deployment to handle new terrain.

Walk These Ways is a quadruped locomotion method from Gabriel Margolis and Pulkit Agrawal at MIT's Improbable AI Lab, posted to arXiv in December 2022 and an oral-presentation paper at CoRL 2022. When a reinforcement-learning-trained walking policy fails outside its training distribution, the usual fix is to go back, tweak the reward or environment, and retrain — a slow cycle. The authors instead propose 'Multiplicity of Behavior' (MoB): train a single policy that takes both a velocity command and a set of behavior parameters, including footfall timing (which can switch between trotting, bounding, pacing, and other gaits), stepping frequency, body height, body pitch, stance width, and foot-lift height. Different parameter combinations complete the same task in different ways with different generalization properties, and a person or a script can switch between them live at deployment without retraining. The policy is trained in Isaac Gym and deployed on the Unitree Go1 at a 50 Hz control rate; the code is open-sourced and has often served as a starting point for quadruped locomotion work since.

ExampleThe paper shows that raising the stepping frequency helps when sprinting on slippery ground; climbing stairs works better with a low stepping frequency and high foot lift; and when pushed by a person, lowering the foot lift and widening the stance improves stability.

Also called

Multiplicity of Behavior (MoB) Controller, Walk These Ways: Tuning Robot Control for Generalization with Multiplicity of Behavior

Related

Legged Locomotion · Gait · RL-based Locomotion Control · Isaac Gym · Unitree Go1 · Sim-to-Real Transfer

Sources

Walk These Ways (arXiv 2212.03238)
Walk These Ways 项目主页 (Chinese)
Improbable-AI/walk-these-ways (GitHub)

As of

2022-12