mjlab

What it is and what it does

mjlab is a reinforcement learning and robotics simulation framework that combines Isaac Lab's composable environment API with MuJoCo Warp, a GPU-accelerated physics engine. It provides building blocks for designing robot control tasks with direct access to native MuJoCo data structures, enabling efficient training of policies for humanoid robots and other agents on GPU clusters.

The framework is designed for research and production robot learning workflows. It supports multi-GPU distributed training, motion imitation from reference trajectories, velocity tracking, and policy evaluation. Users define tasks declaratively (e.g., 'Mjlab-Velocity-Flat-Unitree-G1') and train agents using built-in RL algorithms, with integration to Weights & Biases for experiment tracking and checkpoint management.

Use it for:

• Train humanoid robots to follow velocity commands or track reference motions on flat or complex terrain
• Run large-scale parallel simulations across multiple GPUs for faster policy learning
• Evaluate trained policies in simulation before deployment using checkpoint loading from experiment tracking
• Prototype new robot control tasks using composable environment building blocks and native MuJoCo data access
• Conduct robotics research with reproducible, GPU-accelerated physics simulation and integrated logging

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

mjlab provides a GPU-accelerated reinforcement learning and robotics simulation framework combining Isaac Lab's manager-based API with MuJoCo Warp for composable environment design and training.

Yes, if you are conducting GPU-accelerated robot learning research or building production robot control systems. The active maintenance, production-stable status, permissive license, and integration with standard ML tools (PyTorch, Weights & Biases, TensorBoard) make it a solid choice. The large dependency footprint and GPU requirement are expected for this use case. Not suitable for CPU-only or lightweight simulation needs.

Install

mjlab on PyPI

pip install mjlab

uv add mjlab

poetry add mjlab

Installing mjlab

Before you install

Low install friction with a pure-Python wheel. Active maintenance (last commit 2026-08-14, 5 days old) and production-stable status. However, 20 runtime dependencies including torch, mujoco, and warp-lang create a substantial dependency footprint; installation will pull in significant machine-learning and simulation libraries.

License in practice

Apache-2.0 permissive license allows commercial and private use. Some utilities are forked from NVIDIA Isaac Lab (BSD-3-Clause); those components retain their original licenses per file headers.

Quickstart

# Install
pip install mjlab

# Run demo (requires NVIDIA GPU for training)
from mjlab import demo

# Or use command-line interface
# uv run train Mjlab-Velocity-Flat-Unitree-G1 --env.scene.num-envs 4096

NVIDIA GPU required for training; macOS supported for evaluation only. Requires Python 3.10–3.13.

Verify before relying

• Whether torch and mujoco-warp installation on macOS works for evaluation-only workflows despite GPU-training requirement
• Specific CUDA version compatibility requirements not stated in fact sheet
• Whether all 20 runtime dependencies are strictly required or if some are optional for basic usage

Package facts

Evidence: mjlab-1.6.0-py3-none-any.whl

Keywords: mujoco, mujoco-warp, simulation, reinforcement-learning, robotics

Tags

More Scientific/Engineering packages

Further reading