Your manipulation policy just got legs. Literally.

What it does UMI on Legs is a framework for running manipulation skills on quadruped robots with arms. It trains a whole-body controller in simulation using reinforcement learning, then deploys it in the real world alongside existing visuomotor policies. The idea is to keep the arm doing what it already knows while the legs adjust to maintain balance and positioning.

The interesting bit The system treats the arm’s visuomotor policy as a black-box trajectory generator and wraps it with a learned whole-body controller that handles balance, gait, and base pose. It is essentially a compatibility layer between a fixed arm brain and a new set of legs, trained entirely in simulation before sim-to-real transfer.

Key highlights

• Plug-and-play support for existing visuomotor policies on quadruped hardware.
• Simulation training pipeline for whole-body control, built on IsaacGym and PPO.
• Real-world deployment stack including an iPhone odometry app, ARX5 arm SDK, and hardware assembly guides.
• Modular codebase published as separate submodules meant to be cherry-picked.
• Includes starter checkpoints and evaluation data for immediate simulation rollout.

Caveats

• A TODO comment in the README notes that the UMI code itself may not be present yet, which suggests some components are still incomplete.
• The repository is organized as a loose collection of submodules, so finding the right entry point requires reading the docs carefully.

Verdict Worth exploring if you are building mobile manipulation systems and want to reuse existing arm policies without retraining them. Skip it if you need a fully integrated, turn-key codebase without hardware assembly or simulation tuning.