Open Source RoboticsRobots, software & learning

mjlab

Software

Manager-based robot-learning framework combining Isaac Lab-style environment APIs with GPU-accelerated MuJoCo Warp. [1]

Installation & first result →Official websiteGetting started

For
Researchers building GPU robot-learning environments [1]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
operating system not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
Apache-2.0 · license not fully assessed for model weights, datasets · Details by part

Requirements · Research details and dates

On this page

Overview

mjlab combines a manager-based environment API inspired by Isaac Lab with MuJoCo Warp for GPU robot learning. It includes velocity tracking, motion imitation, training and playback workflows. Its documented platform boundary is useful: training needs Linux and an NVIDIA GPU, while evaluation also supports macOS and Windows through WSL. [1] [4]

Is this for you?

Known limitations
  • The demo requires separately downloaded checkpoint and motion assets.
  • A successful process return does not alone prove the demo ran: download RuntimeError is caught and returns.
  • Dependency API resemblance does not prove Isaac Lab tasks are drop-in compatible.

Editorial assessment from cited documentation; no runtime testing.

[1] [3] · Editorial assessment

What you need

Dependencies
  • Python >=3.10,<3.14
  • MuJoCo and MuJoCo Warp ~=3.11.0
  • PyTorch >=2.7.0
  • Warp >=1.14.0
  • rsl-rl-lib 5.4.2

[2]

Compute for the workload
  • Training: Linux and NVIDIA GPU; CUDA 12.4+ recommended.
  • Evaluation: Linux, macOS or Windows through WSL according to upstream.

[4]

Installation & first result

Open the upstream getting-started guide [4]

  • Check out v1.6.0 with a supported Python version.
  • Run uv run demo with internet access for checkpoint and motion downloads.
  • Verify the viewer shows motion; the demo can return after an asset-download error.

Editorial assessment from cited documentation; no runtime testing.

[1] [3] · Editorial assessment

Installation

Check out v1.6.0 and run uv run demo to provision the source environment and start the pretrained example.

Editorial assessment from cited documentation; no runtime testing.

[1] [2] · Editorial assessment

Hardware, models & data

Inputs and outputs

Configured simulation tasks and policies generate rollouts; training produces checkpoints, while the demo plays a downloaded tracking policy.

[1] [3]

Project notes

Try a pinned demo before creating a training task

This review covers v1.6.0, commit 0fb8a681136be94ffc636a3dd423cabb97d91f10. Use Python 3.10–3.13: the package's upper bound is below 3.14, even though the installation guide simply says 3.10 or higher. The release requires MuJoCo/MuJoCo Warp 3.11.x, PyTorch 2.7.0+ and Warp 1.14.0+, and pins rsl-rl-lib to 5.4.2. [2] [4] [6]

Check out that release and run uv run demo. The demo downloads a pretrained checkpoint and motion file, then opens a Viser viewer for the Unitree G1 tracking task, with eight environments by default. Internet access and the external assets are part of this first run; it is not a fully self-contained source checkout. [1] [3]

Verify that the viewer actually displays policy motion. If asset retrieval raises a RuntimeError, the demo prints a failure message and returns before playback; a normal process return alone is therefore insufficient evidence of a successful demonstration. We did not run it for this review. [3]

For training, upstream's flat-ground example is uv run train Mjlab-Velocity-Flat-Unitree-G1 --env.scene.num-envs 4096. Treat 4096 as an example configuration to size for available GPU memory, not a promised capacity on every device. Motion imitation additionally needs a prepared motion registry or file workflow. [1]

What we checked

Source researched only; no installation, training, benchmark or physical hardware test performed. No affiliate links or reused media.

Not yet established

  • Requirements: operating system, host computer
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Aug 9, 2026 [6]
Latest release
Aug 9, 2026 [6]
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
v1.6.0 / 0fb8a681136be94ffc636a3dd423cabb97d91f10 · software repository

Source files and licenses

Available robot-learning framework with versioned dependencies and clear platform/asset limits.

LayerAvailability / license
softwareavailable · Apache-2.0
Root Apache-2.0; Isaac Lab-derived utilities retain BSD-3-Clause headers. [1] [7]
model weightsavailable · License not fully assessed
Demo downloads external checkpoint/motion; payloads and artifact-specific rights not verified. Availability describes the documented download workflow. [3]
datasetsavailable · License not fully assessed
Demo downloads external checkpoint/motion; payloads and artifact-specific rights not verified. Availability describes the documented download workflow. [3]
documentationavailable · Apache-2.0
Repository materials only; external components retain their own terms. [7]

Keep implementation and asset rights separate

mjlab's root license is Apache-2.0. Its README identifies Isaac Lab-derived utilities that retain BSD-3-Clause headers, and dependencies keep their own terms. Separate demo checkpoint and motion rights were not established here. Similar environment APIs do not by themselves establish that an Isaac Lab task can run unchanged. This profile reports source and documentation findings, not reproduced throughput or policy quality. [1] [7]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history