UW & NVIDIA's SGS lets mega-scale RL solve quadruped locomotion and contact-rich assembly, distilling to real hardware
zhaoran_wang · x · 2026-10-12
- A University of Washington + NVIDIA team (Octi Zhang, Mateo Guaman Castro, Patrick Yin, Byron Boots et al.) introduced Success Guided Sampling (SGS) in a new paper to tackle the exploration bottleneck of large-scale robot RL: an adaptive simulator sampler that concentrates training on task configurations at the policy's capability frontier, instead of uniformly sampling and wasting experience on already-mastered tasks.
- With SGS, RL scales to up to 2^29 (over a billion) parallel locale environments and solves challenging tasks prior methods failed at: agile quadruped locomotion, dexterous contact-rich assembly, and precise whole-body manipulation.
- The team distills the learned manipulation policies into RGB-based policies and demonstrates zero-shot transfer to real hardware on several assembly tasks.
- NVIDIA's SediBY571 shared the paper, noting it further validates the "state-based teacher + distillation into image-based inference policy" recipe for robotics RL, contrasting their own offline-RL approach; built on Isaac Lab and TorchRL.
Related event: UW and NVIDIA Propose SGS for Zero Real-World-Data Robot Training(2 posts)→
More from Embodied
- Voice, facial recognition and animatronic face now run on a Raspberry Pi; beamforming mic array next — natesiggard · 2026-10-12
- RoboWorld Challenge 2026 launches at NeurIPS to test world models beyond visual plausibility — liuziwei7 · 2026-10-12
- Claude recreates a real desk in MuJoCo so an Astra robot arm can paint with variable brush strokes — DJiafei · 2026-10-12
- Cybercab races Waymo on Austin streets as two robotaxis compete in public — Scobleizer · 2026-10-12
- Building a ModRetro game with his daughter and Codex: 5 concrete design fixes — petergyang · 2026-10-12
- Bindu Reddy: superintelligent agent swarms simulating billions of runs will crack mass-market robots — bindureddy · 2026-10-12