NeoteAI and Fudan open 30,000 hours of tactile data for embodied AI
机器之心 · wechat · 2026-07-26
Touch becomes core infrastructure for embodied AI
NeoteAI and Fudan University released a three-part technical report series arguing that touch is no longer just an auxiliary signal for robots. The package includes an open data foundation, a tactile-aware VLA method, and a world model that predicts vision, touch, and actions together.
The data layer
- NeoData contains over 30,000 hours of visual-tactile interaction video.
- It covers about 1.4 million operation clips, 3.3 billion time steps, 8 billion RGB frames, and 10 billion tactile frames.
- The corpus spans 6 robot embodiments, 450 real long-horizon tasks, and data from 90 operators.
- The team also released 5,000 hours of visual-tactile video.
Two model lines
- VTLA predicts the next 50 steps of tactile evolution instead of waiting for delayed touch feedback, helping robots anticipate slip and force changes.
- Reported success rates include 85% on plug insertion and 99% on key pulling, far above vision-only baselines.
- TWAM is a world model with separate video, tactile, and action experts; it reaches 84.5% simulation success and 46.3% on real robots across 8 tasks.
Why it matters
The reports argue that robots need “visual-touch fusion” to handle the last centimeter of physical interaction, and that tactile data may have scaling-law potential similar to vision.
Related event: NeoteAI and Fudan Release Tactile Embodied AI Works(2 posts)→
More from Embodied
- Single-wheel robot demo shows how hard balance control really is — 2C_ornot2C · 2026-07-26
- NeoteAI and Fudan push tactile sensing as a core scaling path for embodied AI — 量子位 · 2026-07-26
- Why robots don’t need human shapes, according to one AI hardware take — VraserX · 2026-07-26
- MicroFactory says its new robots can already assemble electronics — viktor_vrp · 2026-07-26
- Magic Leap exits its own AR headsets and becomes a waveguide supplier — 创业邦 · 2026-07-26
- CVPR 2025 paper learns 3D spatial audio from unlabeled video using camera ego-motion — 量子位 · 2026-07-26