// the find
google-deepmind/mujoco_playground
An open-source library for GPU-accelerated robot learning and sim-to-real transfer.
MuJoCo Playground is a collection of GPU-accelerated RL training environments built on MuJoCo MJX (JAX backend) and now also MuJoCo Warp. It targets researchers who want to train locomotion and manipulation policies fast and then transfer them to real hardware — quadrupeds, humanoids, dexterous hands, and arms like the Panda are all covered out of the box.
The environment coverage is genuinely broad: classic dm_control tasks, seven distinct robot platforms for locomotion, and multiple manipulation setups including dexterous hands. Supporting both MJX and Warp backends gives you a real choice between ecosystem maturity and raw throughput. The domain randomization infrastructure (per-robot `randomize.py` files) is baked in rather than bolted on, which matters for sim-to-real. Colab notebooks for every major environment category lower the barrier to a first run without a GPU workstation.
The README buries a reproducibility gotcha — Ampere GPUs (RTX 30/40 series, which is most researchers' hardware) silently use TF32 by default in JAX, and training results won't match the paper unless you set an env var that isn't set anywhere in the library itself. The training script diverges from the paper's brax training script in undocumented ways (see issue #171), which means you can't fully reproduce published numbers without digging through issues. PyPI is actively discouraged in favor of installing from source, but the source install requires CUDA 12 JAX installed in the right order before syncing, which is the kind of setup that wastes an afternoon. No multi-agent or multi-robot environments — everything here is single-agent.