mjlab is a reinforcement learning and robotics simulation framework that combines Isaac Lab's manager-based API with MuJoCo Warp, a GPU-accelerated physics engine.
The tool addresses the need for efficient, composable environment design in robotics research by pairing Isaac Lab's high-level abstractions with MuJoCo Warp's GPU acceleration. This combination provides direct access to native MuJoCo data structures while maintaining a clean, modular API. The framework requires minimal dependencies and supports multi-GPU training for scaling experiments across hardware.
mjlab suits researchers and practitioners building reinforcement learning agents for robotic control tasks. It is particularly well-suited for projects involving humanoid locomotion, motion imitation, and velocity tracking, as demonstrated by its included training examples. The tool requires an NVIDIA GPU for training, though macOS is supported for evaluation only. The framework's composable building blocks allow users to construct custom environments while retaining the performance benefits of GPU-accelerated physics simulation.
Development activity shows consistent engagement with the research community. The project maintains comprehensive documentation including guides for distributed training and motion imitation preprocessing. The team actively incorporates feedback from collaborators, as evidenced by acknowledgments of feature implementations based on user requests. The codebase includes built-in sanity-check utilities for validating environment designs before training, and the project tracks usage across published research and open-source robotics projects through a dedicated research page.
