mjlab — the framework microduck_rl builds on

microduck_rl is not a training framework built from scratch. It sits on mjlab, a GPU-parallel reinforcement learning environment framework built on MuJoCo Warp.

Knowing about this layer has a practical payoff: the --env.scene.num-envs 4096 argument — simulating 4096 robots at once — is mjlab’s capability, and the CUDA GPU requirement exists because MuJoCo Warp runs on the GPU.

If you want to change the training environment itself rather than only its parameters, this is the repository to read.