real-stanford/umi-on-legs
A robotics framework enabling manipulation policies to run on quadruped robots using whole-body controllers trained with reinforcement learning.

This repository provides a system for combining human demonstrations with simulation-trained whole-body controllers for robot dogs equipped with arms. It includes code for training controllers in simulation, deploying them in real-world scenarios, and integrating existing visuomotor policies onto mobile quadruped platforms. The approach uses manipulation-centric whole-body control to make stationary manipulation skills portable on legged robots.