SuperMimic cuts humanoid training GPU-hours by 1000×: 1 GPU in 3 hours beats SOTA
fdellaert · x · 2026-10-27
SuperMimic claims to scale humanoid control with over 1000× fewer GPU-hours: trained in 3 hours on a single GPU, it surpasses the prior state of the art (128 GPUs for 7 days) across 129,514 motions.
Related work: DMO (CoRL 2025, NYU & LAAS-CNRS) explores a decoupled variant — forward rollouts in a high-fidelity simulator, backward gradients through a learned differentiable dynamics model — avoiding compounding model errors while keeping low-variance analytical gradients. DMO improves sample and wall-clock efficiency on locomotion/manipulation benchmarks and deployed on a Unitree Go2 for quadrupedal and bipedal tasks with robust sim-to-real transfer.
More from Embodied
- Polymarket puts 18% odds on Tesla launching Robotaxis in California this year — Polymarket · 2026-10-27
- Tesla Robotaxi now covers more Texas territory than Waymo, per Polymarket — Polymarket · 2026-10-27
- Hermes Girl Bot returns with dance moves, built on Grok bot + Hermes-Agent — alexcovo_eth · 2026-10-27
- Rust LVGL-style embedded UI library twine ships, built largely with Claude — Necessary-Pilot-3971 · 2026-10-27
- RoboParty Unveils RP1, a Full-Stack Open-Source Bipedal Humanoid at IROS 2026 — emmanuelvivier · 2026-10-27
- NVIDIA open-sources MotionBricks: real-time motion generation at 15,000 FPS with 2ms latency — flngr · 2026-10-27