Isaac Lab's API, freed from NVIDIA's walled garden

What it does

mjlab is a robot learning framework that grafts Isaac Lab’s manager-based environment API onto MuJoCo Warp, DeepMind’s GPU-accelerated physics engine. You write environments the Isaac Lab way, but simulations run on MuJoCo’s data structures instead of NVIDIA’s Isaac Sim. The result is a stack for large-scale RL training with fewer dependencies and no Omniverse-sized installation footprint.

The interesting bit

The project is essentially a compatibility layer with a physics engine swap. That’s the point: Isaac Lab’s API design is genuinely good, but Isaac Sim’s licensing and hardware requirements are a drag. mjlab keeps the abstraction you know while replacing the proprietary backend with an open-source alternative that runs on standard CUDA. It also exposes native MuJoCo data structures directly, so you’re not trapped behind yet another opaque wrapper.

Key highlights

• Multi-GPU distributed training via --gpu-ids flag
• Built-in velocity tracking and motion imitation tasks (Unitree G1 humanoid examples included)
• Dummy agents (zero, random) for MDP sanity checks before real training
• Weights & Biases integration for checkpoint fetching during evaluation
• Apache 2.0 licensed; some utilities forked from Isaac Lab under BSD-3-Clause

Caveats

• Requires NVIDIA GPU for training; macOS limited to evaluation only
• Motion imitation needs preprocessing setup not included in the quick demo
• Relatively young project (2026 paper, ~2.5k stars); ecosystem maturity unclear

Verdict

Worth a look if you’re already productive in Isaac Lab but tired of NVIDIA’s platform requirements, or if you want GPU-parallel MuJoCo with a higher-level API than raw Warp provides. Skip it if you need mature sim-to-real tooling or are committed to Isaac Sim’s rendering and asset pipeline.