Unitree's IsaacLab wrapper: train in sim, walk in reality

What it does

Unitree RL Lab is a training and deployment stack for Unitree’s Go2, H1, and G1-29dof robots. It wraps IsaacLab for GPU-accelerated RL training, then exports policies through a MuJoCo validation layer before they ever touch physical hardware.

The interesting bit

The sim-to-real chain is explicit and non-negotiable: IsaacLab → MuJoCo → physical robot. The README treats MuJoCo not as an alternative simulator but as a required proving ground, complete with elastic-band tricks and joystick rituals to test policy robustness before ethernet-based deployment.

Key highlights

• Supports three current Unitree platforms: Go2 quadruped, H1 humanoid, and G1-29dof humanoid
• Built on IsaacLab 2.3.0 / Isaac Sim 5.1.0 for parallel GPU training
• Provides a unitree_rl_lab.sh helper script that wraps listing, training, and inference tasks
• Includes C++ deployment controllers in deploy/robots/ with unitree_sdk2 dependency
• Dual asset pipeline: USD files from Hugging Face, or URDF via unitree_ros (recommended for IsaacSim ≥ 5.0)

Caveats

• Asset setup is manual: you must download robot models separately and edit hardcoded paths in unitree.py
• Deployment section is G1-29dof-centric; other robots are mentioned but not documented for sim2real steps
• “on-borad control program” typo in the sim2real warning suggests the docs haven’t been heavily polished

Verdict

Grab this if you own or lease Unitree hardware and want a vendor-supported IsaacLab starting point. Skip it if you’re looking for a general humanoid RL framework — the Unitree-specific asset plumbing and deployment code make it a branded toolchain, not a research substrate.