SuperMimic trains humanoid control on 129k motions with 1 GPU in 3 hours, cutting GPU-hours 1000x
kevin_zakka · x · 2026-10-22
qiayuanliao's team released SuperMimic, a scalable approach to humanoid control that combines physics simulators with learned world models: the simulator provides accurate transitions while the world model amortizes them into a much faster, differentiable surrogate for policy optimization, cutting GPU-hours by more than 1000x.
Across 129,514 motions, SuperMimic surpasses the previous state of the art (128 GPUs, 7 days) using a single GPU in 3 hours. Kevin Zakka calls it a nice example of physics sims and learned world models working together, with a browser-playable demo available.
More from Embodied
- Pliny to jailbreak a Unitree robot live on stage at Black Hat EU — Kyrannio · 2026-10-22
- Agility Robotics co-founder Jonathan Hurst to host AMA on Digit 5 robot — ___Mufasaa · 2026-10-22
- Why a 2020 Paper on Rotation Representations Is Called the Biggest Kinematics Breakthrough in a Decade — andrew_n_carr · 2026-10-22
- USTC demos SpiRobs, an octopus-inspired logarithmic spiral robot — burny_tech · 2026-10-22
- Hugging Face's Microducks robot tops 15,000 units sold, closes out Nvidia's European tour — huggingface · 2026-10-22
- FunCAD generates working, 3D-printable CAD mechanisms from a single text prompt — jonLorraine9 · 2026-10-22