mujocolab/mjlab

What it solves

mjlab provides a lightweight, GPU-accelerated framework for robot learning. It addresses the need for a high-performance simulation environment that combines the ease of use of manager-based APIs (like Isaac Lab) with the GPU acceleration of MuJoCo Warp.

How it works

The framework integrates the manager-based API design from Isaac Lab with MuJoCo Warp, a GPU-accelerated version of the MuJoCo physics engine. This allows users to create composable building blocks for environment design while maintaining direct access to native MuJoCo data structures with minimal dependencies.

Who it’s for

Robotics researchers and developers who need to train agents (such as humanoid robots like the Unitree G1) for tasks like velocity tracking and motion imitation using GPU-accelerated simulation.

Highlights

• GPU Acceleration: Leverages MuJoCo Warp to enable high-speed training on NVIDIA GPUs.
• Composable Design: Uses a manager-based API for flexible environment design.
• Scalability: Supports multi-GPU training to scale the number of environments.
• Detailed Examples: Includes built-in support for tasks like velocity tracking and motion imitation for humanoid robots.