Learning Quadruped Walking from Seconds of Demonstration

TL;DR

This paper demonstrates that quadruped locomotion can be effectively learned from just seconds of demonstration by leveraging the structure of limit cycles and neural network properties, enabling robust policies trained offline.

Contribution

It introduces a new imitation learning method that aligns latent space variations with output actions, explaining why small data regimes are effective for quadruped locomotion.

Findings

A few seconds of demonstration suffice for training robust locomotion policies.

The method outperforms traditional approaches in data efficiency and robustness.

Hardware experiments validate the effectiveness of the approach.

Abstract

Quadruped locomotion provides a natural setting for understanding when model-free learning can outperform model-based control design, by exploiting data patterns to bypass the difficulty of optimizing over discrete contacts and the combinatorial explosion of mode changes. We give a principled analysis of why imitation learning with quadrupeds can be inherently effective in a small data regime, based on the structure of its limit cycles, Poincar\'e return maps, and local numerical properties of neural networks. The understanding motivates a new imitation learning method that regulates the alignment between variations in a latent space and those over the output actions. Hardware experiments confirm that a few seconds of demonstration is sufficient to train various locomotion policies from scratch entirely offline with reasonable robustness.