Touch-R1 uses RL for cross-sensor tactile reasoning, beating GPT-4o by 24.7 points
机器之心 · wechat · 2026-11-09
Damo-style robotics startup DawnVision, with Xiamen/Nanjing University and others, proposed Touch-R1 — the first model to systematically apply reinforcement learning to cross-sensor tactile reasoning, accepted to NeurIPS 2026.
Problem: when vision and touch conflict (looks like hard plastic, feels soft), GPT-4o and Gemini-2.5-Pro tend to trust vision, risking grasp failures and safety issues.
Infrastructure
- TouchReason-1M dataset: 1,000+ objects, 9 material classes, 4 visuotactile sensors, standardized actions at 20N force; 14,700 valid tactile sequences and 1.32M synchronized data pairs.
- TouchReason-Bench: 4,800 QA pairs covering physical property recognition, ordinal reasoning, cross-sensor consistency.
Method: three-stage training — future-frame prediction cold start, structured-QA SFT forming a "perceive-compare-conclude" chain, tactile-tailored GRPO RL with cross-sensor consistency and ordinal rewards to prevent guessing right without truly understanding touch.
Results: Touch-R1-7B leads on all 9 core metrics with a composite score of 60.1 — +18.4 over tactile specialist Octopi-13B, +24.7 over GPT-4o, +12.6 over prior best SToLa — and can overturn visual misjudgment using tactile evidence.
More from Embodied
- TacClaw-Infinity: 10,000-Hour Vision-Tactile Dataset Lifts π0.5 Success from 50% to 74% — CyberRobooo · 2026-11-09
- Waymo explains at Hot Chips why camera-only autonomy plateaus at 99.9% — 4310sy · 2026-11-09
- 'The future is here, just unevenly distributed': humanoid robot spotted at a conference — technollama · 2026-11-09
- OriginFlow CEO on physical AI data: real moats are systems engineering and economic efficiency, not data volume — 创业邦 · 2026-11-09
- Ex-Zhiyuan PM's embodied data startup tops $3M in orders; bottleneck is data processing, not collection — 快鲤鱼 · 2026-11-09
- Developer builds a physical prompting controller out of OpenAI DevDay merch — paw_lean · 2026-11-09