mjlab
MuJoCo Labmjlab is a GPU-accelerated robot learning framework that pairs Isaac Lab's manager-based API, in which environments are composed from modular observation, reward and event terms, with MuJoCo Warp physics. It ships reference tasks for humanoid velocity tracking, motion imitation and manipulation, batched RGB-D camera, raycast, contact and tactile sensors, and multi-GPU training, and it installs with a single command.
Openness
5 high confidence- license
- Apache-2.0 (OSI, utilities forked from Isaac Lab keep BSD-3-Clause)
- source
- public
- core features withheld
- no — an academic project with no paid offering
mjlab is Apache-2.0 and builds entirely from its public repository, with a few utilities forked from Isaac Lab keeping their BSD-3-Clause terms. It is a research project from its paper's authors with no company or paid edition behind it.
- https://arxiv.org/abs/2601.22074 recorded 2026-11-12
arXiv 2601.22074, "mjlab: A Lightweight Framework for GPU-Accelerated Robot Learning", by Kevin Zakka, Qiayuan Liao, Brent Yi, Louis Le Lay, Koushil Sreenath and Pieter Abbeel: "We present mjlab, a lightweight, open-source framework for robot learning".
- https://raw.githubusercontent.com/mujocolab/mjlab/main/LICENSE recorded 2026-11-12
LICENSE is the Apache License, Version 2.0 text: "Apache License", "Version 2.0, January 2004".
- https://raw.githubusercontent.com/mujocolab/mjlab/main/README.md recorded 2026-11-12
README License section: "mjlab is licensed under the Apache License, Version 2.0"; "src/mjlab/utils/lab_api/ — Utilities forked from NVIDIA Isaac Lab (BSD-3-Clause license, see file headers)"; install from source with git clone and uv run. No paid tier is mentioned.
- https://ungh.cc/repos/mujocolab/mjlab/files/main recorded 2026-11-12
Full listing of the main branch, 624 paths; no ee, enterprise, commercial or proprietary path.
- https://ungh.cc/repos/mujocolab/mjlab/releases/latest recorded 2026-11-12
Latest release mjlab v1.6.0, published 2026-08-09.
Adoption
2 medium confidencePyPI downloads of the mjlab package, which is the framework itself. The README's own quick start clones the repository instead, so source installs go uncounted.
- https://mujocolab.github.io/mjlab/main/source/installation.html recorded 2026-11-12
Installation guide covering installation from PyPI, from source and with Docker.
- https://pypi.org/pypi/mjlab/json recorded 2026-11-12
PyPI metadata for mjlab 1.6.0: license_expression Apache-2.0, Source https://github.com/mujocolab/mjlab.
- https://pypistats.org/api/packages/mjlab/recent recorded 2026-11-12
pypistats: last_month 60,015 downloads for mjlab.
Capability
4 medium confidencemjlab runs thousands of MuJoCo Warp environments in parallel on the GPU and renders camera images for all of them in the same batch. It has one physics engine and a handful of reference tasks, where Isaac Lab offers several physics backends and more than thirty ready-to-train environments, so it sits a step below.
- https://arxiv.org/abs/2601.22074 recorded 2026-11-12
Abstract: mjlab "pairs it with MuJoCo Warp for GPU-accelerated physics" and "ships with reference implementations of velocity tracking, motion imitation, and manipulation tasks".
- https://raw.githubusercontent.com/mujocolab/mjlab/main/docs/source/sensors/rgbd_camera.rst recorded 2026-11-12
Sensor docs: "CameraSensor renders RGB and depth images on the GPU using MuJoCo Warp's ray-traced rendering pipeline"; outputs batched as "[B, 120, 160, 3] uint8".
- https://raw.githubusercontent.com/mujocolab/mjlab/main/README.md recorded 2026-11-12
README: "mjlab combines Isaac Lab's manager-based API with MuJoCo Warp, a GPU-accelerated version of MuJoCo"; training example with "--env.scene.num-envs 4096"; "Multi-GPU Training"; tasks Mjlab-Velocity-Flat-Unitree-G1 and Mjlab-Tracking-Flat-Unitree-G1.
Verified 2026-11-12, except some scores, still awaiting re-confirmation