Zero real-world data: Success-Guided Sampling unlocks zero-shot dexterous robot manipulation via sim2real RL
Scobleizer · x · 2026-10-12
Researchers released Success-Guided Sampling (SGS), a small tweak to the task sampler that, combined with large-scale end-to-end sim2real RL, lets robots mesh gears, thread nuts, and insert pegs zero-shot from RGB with no real-world data or demonstrations. The same recipe transfers to locomotion over much harder terrains. The authors argue the missing piece wasn't a better RL algorithm but changing what the robot practices — just PPO at scale, no per-task reward tuning. NVIDIA's gavrielstate amplified the work, praising the value of scaled RL training in simulation and teasing 'even cooler' follow-ups. All demo videos run at 1x speed.
Related event: UW and NVIDIA Propose SGS for Zero Real-World-Data Robot Training(2 posts)→
More from Embodied
- Bindu Reddy: superintelligent agent swarms simulating billions of runs will crack mass-market robots — bindureddy · 2026-10-12
- EgoLAP paper adds motion-level language reasoning to VLA pretraining, lifting real-world robot task progress to 80.1% — Majumdar_Ani · 2026-10-12
- Vernata: ETH Zurich's multi-teacher distillation for self-supervised LiDAR representations, IROS 2026 — rsasaki0109 · 2026-10-12
- Dev asks Grok to build itself a Vision Pro app, creates conversational room bots — Baconbrix · 2026-10-12
- Local AI RPG stack combines 3 DGX Sparks, GLM 5.3 Flash and ComfyUI — HankYeomans · 2026-10-12
- AI Agent Does CAD: KiCad MCP Reads Mounting Holes for a Robot Build — burhop · 2026-10-12