MacMimic trains humanoid motion controllers in 22 minutes on a MacBook Pro

KyleMorgenstein · x · 2026-11-20

MacMimic is an open-source (MIT) project that trains a simulated Unitree G1 humanoid to track 40 mocap clips — walking, sprinting, dancing, getting up, cartwheels — in just 22 minutes on a MacBook Pro (M4 Pro), no CUDA or cloud needed. MuJoCo simulates 4,096 robots on CPU cores while MLX trains on Apple GPU; each iteration takes 0.9s using a model-based approach (learn robot physics, then differentiate through it). It hits the benchmark target 2.04x faster than a BlinkMimic MLX port and 2.6x faster than XMimic. Full code and training pipeline are public, showing RL training for humanoid control can now run on consumer laptops.

Original post →

More from Embodied

Embodied channel →