Research
Open source humanoid motion trackers now train in minutes on one GPU after Berkeley's SuperMimic
Training a controller that lets a humanoid follow human motion clips used to take a GPU cluster. In under four weeks that has dropped to a few minutes on a single card, Humanoids Daily reports, thanks to a UC Berkeley paper and a run of open source projects built on top of it.
The paper, SuperMimic, was posted to arXiv on 20 October by a team including Pieter Abbeel and Koushil Sreenath. Instead of the usual model-free reinforcement learning, it trains a neural network to predict how the robot's body responds to actions, then improves the tracking policy by backpropagating through that learned model. The authors say three hours on one GPU beat a baseline trained for seven days on 128 GPUs, and that the policy transferred to a physical Unitree G1.
The paper shipped without code. An ETH Zürich researcher published an unofficial reimplementation the next day, and others followed: InstantTrack, BlinkMimic and XMimic each took the previous project's code and made it faster, Humanoids Daily says. The official Berkeley code appeared on GitHub around 10 November.
BlinkMimic, whose paper was posted on 12 November, reports reaching its target in 3.4 minutes on one RTX 4090, and XMimic reports 0.86 minutes on four H100s, both on a 40-clip benchmark in simulation. A community list, awesome-supermimic, now tracks more than 40 implementations and extensions.
The caveat is important: none of the community projects reports a test on a real robot, and their speed figures come from a narrow benchmark of 40 motion clips.
