mjlab

Isaac Lab API on top of GPU-accelerated MuJoCo Warp for RL and robotics research

License

Apache-2.0

Overview

mjlab combines Isaac Lab's manager-based environment API with MuJoCo Warp, the GPU-accelerated version of MuJoCo. The framework provides composable building blocks for environment design with minimal dependencies and direct access to native MuJoCo data structures, targeting reinforcement learning and robotics research. Maintained by the mujocolab organization with active GitHub Actions CI, nightly benchmarks, PyPI distribution, and growing model zoo for humanoid (Unitree G1, Booster T1) and quadruped (Unitree Go1) locomotion.

mjlab is a research framework that wraps MuJoCo Warp -- Google DeepMind's GPU-accelerated MuJoCo backend -- in the manager-based environment API popularised by NVIDIA Isaac Lab. The result is a single Python library where researchers can describe a robot, scene, observations, rewards, randomisations, and curricula as composable managers and then train policies at thousands of parallel environments per GPU.

The project ships ready-to-train tasks for humanoid velocity tracking (Unitree G1, Booster T1) and quadruped locomotion (Unitree Go1) on flat and rough terrain, plus motion-tracking from CMU MoCap data. Distributed training across multiple GPUs is supported out of the box, evaluation can run while training (pulling the latest checkpoint from Weights & Biases), and policies can be exported and replayed in stand-alone MuJoCo sims.

Unlike Isaac Lab proper -- which depends on Isaac Sim and the NVIDIA RTX stack -- mjlab keeps dependencies minimal and exposes native MuJoCo data structures directly, making it easier to plug into existing MuJoCo-based research code and to run on any CUDA GPU. It is published on PyPI, has nightly benchmarks, and is actively maintained by the mujocolab organisation.

Why it matters

Isaac Lab's manager-based API is now the de-facto standard for large-scale parallel RL on robotics tasks, but it's tied to NVIDIA Isaac Sim. mjlab brings the same ergonomic API to the open-source MuJoCo ecosystem, removing the Isaac Sim lock-in and making fast GPU-parallel RL accessible to any team already invested in MuJoCo or running on non-RTX hardware.

Problem it solves

Researchers wanting Isaac Lab's clean manager-based API have had to commit to NVIDIA Isaac Sim, which is heavyweight, RTX-locked, and far from the existing MuJoCo research ecosystem. mjlab gives those same ergonomics on top of MuJoCo Warp, so teams already comfortable with MuJoCo (or running on commodity CUDA GPUs) can adopt the same authoring patterns without switching simulators.

Key takeaways

• Reuses Isaac Lab's manager-based environment API but runs on GPU-accelerated MuJoCo Warp instead of Isaac Sim
• Ships training tasks for Unitree G1 / Booster T1 humanoids and Unitree Go1 quadruped, plus CMU MoCap motion tracking
• Native multi-GPU training, zero-install demo via uvx, PyPI package, Google Colab notebook
• Apache-2.0 licensed, actively maintained (2,200+ stars, nightly benchmarks, CI)
• Direct access to MuJoCo data structures keeps integration with existing MuJoCo research code straightforward

Who it is for

• RL researchers training humanoid and quadruped locomotion at scale
• Robotics labs already invested in MuJoCo who want Isaac-Lab-style ergonomics
• Sim-to-real teams needing lightweight, scriptable simulation with CMU MoCap-driven motion tracking
• Engineers running on non-RTX CUDA hardware who can't use Isaac Sim

Limitations

• Requires an NVIDIA GPU for training; macOS is evaluation-only
• Younger project than Isaac Lab itself -- ecosystem of pre-built tasks is still small (humanoid + quadruped locomotion focus)
• Tied to MuJoCo Warp, which is itself relatively new and has its own evolving feature gaps vs. classic MuJoCo
• No built-in vision/perception stack -- pairs best with policies that consume proprioceptive state