UW team's Success-Guided Sampling scales RL to 1M simulated robots for contact-rich manipulation
KyleMorgenstein · x · 2026-11-09
A University of Washington + NVIDIA team (CoRL 2026) released Success-Guided Sampling (SGS), addressing the exploration bottleneck in mega-scale RL by training only on task configurations neither too easy nor too hard for the current policy — a modification to standard PPO.
Key results:
- Keeps improving as parallel environments scale to 1 million simulated robots, where uniform sampling falls behind.
- Locomotion: a single from-scratch policy takes ANYmal C/D across stepping stones, floating islands, climbing boxes and more.
- Manipulation: UR5e and Franka learn contact-rich NIST task board assembly with zero demonstrations and one shared reward function.
- Sim-to-real: simulation-only UR5e assembles real parts zero-shot from RGB images — no demos, retargeting, or tactile sensors.
- Code coming soon; presenting at CoRL Poster Session #1.
More from Embodied
- Embedded Engineer's Essay: Bolting an LLM Onto a Device Isn't AI-Native Hardware — LocksmithOk2525 · 2026-11-09
- CoRL 2026 lineup: Fei-Fei Li on Atlas world model, fireside chat with Lisa Su — PeterStone_TX · 2026-11-08
- MoCap-to-video pipeline: RGBMoCap + Blender + Seedance 2.5 re-renders K-pop choreography — Choidonhyeon · 2026-11-08
- PrimeBOT ships 10,000th personal humanoid, doubling output in 14 days — CyberRobooo · 2026-11-08
- Apple Reportedly Prepping a $1,000 Home Robot for 2027, Handling FaceTime and Lights — emmanuelvivier · 2026-11-08
- OpenRoboKit launches open-source robotics toolkit series; RoboKit paper wins CoRL 2026 Spotlight — CyberRobooo · 2026-11-08