NeoteAI open-sources 5,000 hours of real tactile data plus three tactile foundation reports
青稞AI · wechat · 2026-11-19
NeoteAI, with Fudan University's Institute of Trustworthy Embodied AI, released a tactile research series spanning foundation representations, manipulation policies, and world models, alongside 5,000 hours of open-sourced real vision-tactile interaction data.
- Data & representation: NeoData holds 30k+ hours of visuo-tactile data, 1.4M manipulation episodes, 3.3B timesteps from 90 operators across 6 robot platforms and 450 long-horizon tasks; NeoForce aligns heterogeneous sensor signals (gel, capacitive, 6-axis force) into standardized 3-axis force fields. Open data has 500k+ ModelScope and 200k+ Hugging Face downloads.
- VTLA policy: predicts tactile evolution 50 steps ahead; 85% plug insertion, 99% key extraction, 63.8% average on 20 sim tasks; long-horizon tasks (towel folding etc.) reach 95%/80%/75%.
- TWAM world model: jointly predicts future vision, touch, and action; 84.5% average sim success, 46.3% on 8 real-robot tasks.
Industry context: 40+ disclosed funding rounds totaling 5B+ RMB in China's tactile embodied space in Jan–Oct 2026 alone.
More from Embodied
- Robotics sim tips: calibrate your real robot first, never hand-tune sim parameters — jonstephens85 · 2026-11-19
- Embedl cuts Jetson Thor driving-plan latency from 603.7ms to 82.7ms without touching weights — AIFlow_ML · 2026-11-19
- BiCAT builds bimanual coordination directly into VLA action tokens — DJiafei · 2026-11-19
- NVIDIA ships Alpamayo 2 Edge: 2B reasoning VLA model running in-vehicle at 160-190ms latency — tomaszbednarz · 2026-11-19
- Game2Player Turns 568 Web Games Into Verifiable Training Grounds for Visual Agents — UIUC-CS · 2026-11-19
- Musk predicts Tesla Optimus will wash, dry and put away dishes in 12-18 months — Polymarket · 2026-11-19