MSD: Learning general and agile locomotion using multi-skill distillation

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Legged robots are particularly well-suited for navigating terrains that are inaccessible to wheeled robots, which has led to a significant increase in their applications in recent years. However, many existing control approaches still struggle to achieve generalization in diverse and unstructured environments. In this work, we introduce the MSD training framework for general and agile locomotion of legged robots through a multi-skill distillation approach. First, expert policies are trained for different terrains. Then, using the DAgger algorithm, these expert skills are distilled into a single deployable policy. The proposed policy integrates depth images as exteroceptive observations, enabling agile navigation across diverse and unstructured terrains. Finally, deployment on the MagicDog-W robot validates the policy’s capability to achieve robust and flexible locomotion in challenging environments.

This publication has 3 references indexed in Scilit: