mjlab

A lightweight GPU robot-learning framework combining an Isaac Lab-style interface with MuJoCo Warp physics.

mjlab is an open-source framework developed by Kevin Zakka, Koushil Sreenath, Pieter Abbeel, and colleagues, with its paper released in January 2026. It reuses Isaac Lab's “manager-based” interface, where observations, rewards, randomization events, and so on are written as composable modules to assemble an environment; its physics backend is instead MuJoCo Warp, the GPU version of MuJoCo, which can simulate thousands of environments in parallel. Its advantages are that it installs with a single command, has few dependencies, and gives direct access to MuJoCo's native data structures. It ships with three reference task categories — velocity tracking, motion imitation, and manipulation — commonly used for humanoid and quadruped reinforcement-learning locomotion control. Training requires an NVIDIA GPU; macOS can only be used for evaluation.

ExampleRunning uv run train Mjlab-Velocity-Flat-Unitree-G1 --env.scene.num-envs 4096 trains a Unitree G1 to walk according to velocity commands across 4,096 parallel environments.

Related

MuJoCo Warp · NVIDIA Isaac Lab · MuJoCo Playground · GPU-Accelerated Parallel Simulation · RL-based Locomotion Control · Unitree G1

Sources

mjlab: A Lightweight Framework for GPU-Accelerated Robot Learning (arXiv 2601.22074)
mjlab GitHub 仓库 (Chinese)

As of

2026-09