Robots can see, but useful manipulation often fails at the fingertips. NeoteAI and Fudan University's TEAI team released three N0 technical reports: N0-Foundation, N0-VTLA and N0-TWAM. The package includes research pages, reports, data, code and checkpoints.
Touch Becomes A Data Layer
N0-Foundation reports more than 30,000 hours of visual-tactile interaction data, 1.4 million episodes, 3.3 billion timesteps, 8 billion RGB frames and 10 billion tactile frames. The corpus spans six embodiments and more than 450 tasks.
The open subset is smaller but still material: 5,000 hours across six embodiments, over 250 tasks and more than 200 skills. That makes the release useful, while keeping the important caveat clear: most of the full corpus is still not public.
The Model Predicts Future Contact
N0-VTLA does not simply concatenate touch images with camera frames. It compresses current contact into latent tactile tokens, then predicts how touch will change over the next action chunk. In plain terms, the robot tries to anticipate slipping, jamming or seating before the next movement is completed.
The team reports 47.2% average success on nine NeoReal real-robot tasks, compared with 29.4% for the pi0.5 baseline. Socket plugging reaches 85% against 60%, and board insertion moves from zero for both baselines to 25%.
A World Model That Includes Feeling
N0-TWAM goes further by generating future video, future touch and action inside one world-action model. The model has 7.16B trainable parameters and was pretrained for 30,000 steps on 128 H800 GPUs.
Its reported average success is 84.5% on UniVTAC, 49.4% on NeoSim and 46.3% on eight real-robot tasks. Those are team-reported benchmark settings, so the next checks are third-party reproduction, outside submissions to NeoReal or NeoSim, and whether tactile sensors can be manufactured and calibrated reliably.
Sources: QbitAI, NeoteAI/Fudan N0-Foundation technical report, N0-VTLA technical report, N0-TWAM technical report, Hugging Face dataset, GitHub code repositories, CocoLoop; verified 30,000 interaction hours, 1.4 million episodes, 3.3 billion timesteps, 8 billion RGB frames, 10 billion tactile frames, 5,000-hour open subset, NeoReal success rates, N0-TWAM 7.16B parameters, 128 H800 GPUs and benchmark scopes.