Robots can see a cup. They still crush it. A 30,000-hour open dataset of touch is trying to fix the 'last centimeter.'

A wall outlet sits at a robot's-eye height, and the arm has clearly identified it. The prongs touch plastic. They skate. They bump. They miss. Vision identified the outlet from across the room, but vision does not know that the socket is recessed, that the cord is stiff, or that the prongs have already begun to bend. The contact moment is where most home and factory robots still fail, and it is the moment three open-source papers from Shanghai-based NeoteAI and Fudan University's Trusted Embodied Intelligence Research Institute try to fix.

The team released a coordinated trio, N0-Foundation, N0-VTLA, and N0-TWAM, on July 26, with data, code, and model weights public at research.neoteai.com. The frame is direct: tactile sensing, the sense of pressure, slip, and texture, is the missing modality for embodied AI, the branch of AI that drives physical robots. Where language models scaled on more text and image models scaled on more pixels, the bottleneck for robots is hours of contact.

N0-Foundation is the substrate. It introduces NeoData: 30,000+ hours of visual-tactile interaction across six robot embodiments including Franka, Piper, and UR5e, with roughly 1.4 million manipulation clips, 3.3 billion timesteps, 8 billion camera frames, and 10 billion tactile frames. Ninety human operators collected 450 long-horizon tasks, of which 5,000 hours are open-sourced now. The companion model, NeoForce, is a unified representation that fuses readings across different tactile sensor types, a known fragmentation problem, so downstream code can treat "slip" the same way regardless of which glove or skin sits on the gripper.

N0-VTLA is where the contact numbers land. It injects 50-step tactile prediction directly into a VLA (vision-language-action) policy, the dominant family of robot control models. The team's argument is that current VLA systems act like a driver using only the rearview mirror: they see the result of contact after it happens and then try to recover. Plug-insertion success rises from 60% with vision-only inputs to 85% with the tactile-injected VLA. Key-removal, pulling a key from a tight lock, moves from 35% to 99%. On long-horizon household skills, the same system, trained with offline reinforcement learning, lifts towel-folding from 50% to 95%, bag organization from 35% to 80%, and box folding from 20% to 75%.

N0-TWAM is the imaginative piece. A world model is a simulator that predicts what a scene will look like next; N0-TWAM is a tactile-aware world model that also predicts what the robot will feel. It uses a mixture-of-experts design with three asynchronous experts for video, tactile, and action streams, totaling 7.2 billion parameters, about half the cost of a full-width variant. In a simulation benchmark, it hits 84.5% task success versus 36% for the strongest world-model baseline. On a real-world eight-task average, it reaches 46.3% versus 21.9% for LingBot-VA, a recent open baseline. Ablations show both future-tactile and current-tactile conditioning matter: remove either and performance drops.

The substrate matters more than any single number. Vision-first robotics is a closed-loop story: more cameras, better annotations, bigger VLAs. Touch-first robotics has been a patchwork; one lab's tactile glove does not speak another's taxel skin, and few datasets scale past a few hundred hours. The N0 release is the first public attempt, at this scale, to put tactile on the same kind of open, versioned, downloadable footing that image data has had since ImageNet. QbitAI's coverage, Sina Finance's report, and NetEase's write-up all carry the same figures from NeoteAI's materials, so the release is the news.

Three caveats run through the NeoteAI materials. Real-world data collection is expensive; the 30,000 hours took 90 operators. Cross-embodiment generalization is partly answered, partly open: does a model trained on Franka and Piper transfer to a new arm? Real-time inference on the tactile-aware world model at 50 Hz is an engineering question the papers do not settle. And general embodied intelligence, the kind that lets a robot unpack a new kitchen without a tutorial, remains far off. NeoteAI, founded in 2025 in Shanghai, closed an angel round of roughly 100M RMB (about $14M USD at mid-2026 rates) in May 2026 led by the Fudan Innovation Centre and Shanghai STVC Group, so the release is foundation-building, not product delivery.

The next watch item is independent reproduction. The 85% and 99% numbers come from NeoteAI's own evaluation harness. If a second lab replicates the plug-insertion gap on a different arm by the end of the year, the tactile axis is real. If not, the contact moment is still waiting.