mjlab
mjlab is a robot learning framework that combines Isaac Lab’s manager-based API with MuJoCo Warp, a GPU-accelerated version of the MuJoCo physics simulator. It provides building blocks for designing environments and training policies, with minimal dependencies and direct access to MuJoCo’s data structures.
mjlab needs an NVIDIA GPU for training. macOS is supported for evaluation only.
Try it
You can try the mjlab demo without installing anything:
uvx --from mjlab --refresh demo
Or run it in Google Colab, with no local setup.
Example: Unitree RL Mjlab
Unitree RL Mjlab is Unitree’s reinforcement learning project built on mjlab, supporting robots such as the Unitree G1 and H1_2. It’s a useful example of the whole workflow, from training a policy in simulation to deploying it on a real robot:
Train → Play (replay the trained policy in simulation) → Sim2Real (deploy it on the real robot)
MuJoCo
mjlab is built on MuJoCo, an open-source physics engine widely used in robotics research.