mjlab: A Lightweight Framework for GPU-Accelerated Robot Learning

TL;DR

mjlab presents a lightweight, GPU-accelerated framework for robot learning built on MuJoCo Warp and a manager-based API, prioritizing fast iteration, transparent physics, and MuJoCo-native workflows. By exposing MuJoCo data structures directly and providing a PyTorch-ready training interface, it enables thousands of parallel environments with minimal setup and dependency friction. The framework ships with three robot morphologies and three reference tasks to demonstrate locomotion, motion imitation, and manipulation within a modular, extensible architecture. Its CLI-first configuration, collocated definitions, and strong typing support AI-assisted development and broad community contributions, aiming to accelerate sim-to-real research and scalable multi-robot experimentation.

Abstract

We present mjlab, a lightweight, open-source framework for robot learning that combines GPU-accelerated simulation with composable environments and minimal setup friction. mjlab adopts the manager-based API introduced by Isaac Lab, where users compose modular building blocks for observations, rewards, and events, and pairs it with MuJoCo Warp for GPU-accelerated physics. The result is a framework installable with a single command, requiring minimal dependencies, and providing direct access to native MuJoCo data structures. mjlab ships with reference implementations of velocity tracking, motion imitation, and manipulation tasks.

Figures (4)

• Figure 1: Unitree G1 humanoids performing a dance trained with the motion tracking pipeline from BeyondMimic liao2025beyondmimic. Thousands of parallel environments are simulated on a single GPU and visualized in mjlab's web-based Viser viewer.
• Figure 2: Entities are composed into an MjSpec, compiled, and transferred to MuJoCo Warp for GPU simulation. The ManagerBasedRlEnv orchestrates the MDP; RSL-RL handles training.
• Figure 3: A composite terrain grid generated by mjlab. Sub-terrain patches include flat ground, stairs, sloped pyramids, heightfield noise, and wave patterns. Difficulty increases along one axis for curriculum-based training.
• Figure 4: The three robot morphologies shipped with mjlab: Unitree G1 humanoid (left), Unitree Go1 quadruped (center), and YAM robot arm (right). All models are adapted from MuJoCo Menagerie zakka2022menagerie.