SuperMimic cuts humanoid training GPU-hours 1000×; differentiable physics trains 512-robot swarms
zhaoran_wang · x · 2026-10-22
Two major efficiency results in robot learning:
- SuperMimic (Qiayuan Liao et al.): humanoid whole-body control across 129,514 motions — 1 GPU for 3 hours beats prior SOTA (128 GPUs, 7 days), a >1000× reduction in GPU-hours.
- Quadruped swarm paper (arXiv 2606.23153, Yuang Zhang, Yunlong Song et al.): "asymmetric physics" end-to-end learning — a high-fidelity non-differentiable simulator provides realistic motion/contact dynamics while differentiable surrogate models supply gradients, letting up to 512 quadrupeds learn decentralized vision-based navigation in obstacle-rich environments. 50× better sample efficiency than PPO; deployment uses only a single forward depth camera with no explicit communication, centralized planning, or global maps. Policies zero-shot transfer to six real quadrupeds across five scenarios, showing predictive avoidance and right-side yielding.
Yunlong Song suggests extending differentiable physics to perceptive whole-body control and locomotion as a free research idea.
Related event: SuperMimic Trains Humanoid Whole-Body Control 1000x Cheaper on a Single GPU(3 posts)→
More from Embodied
- LAGEN: one frame of latency can cut real-time agent performance by 95% — wzenus · 2026-10-22
- Oopsie Data: multi-lab effort to build first large-scale robot failure dataset — micoolcho · 2026-10-22
- Humanoid robot fights spark debate: entertainment is the only real use case today — KyleMorgenstein · 2026-10-22
- Unitree H2X humanoid robot sports dual battery slots in hands-on shots — micoolcho · 2026-10-22
- Unitree developer day draws 3,000 attendees as its robotics ecosystem takes shape — micoolcho · 2026-10-22
- Unitree's New H2-X Humanoid Pictured Next to G1, Height Gap Revealed — micoolcho · 2026-10-22