N0-VTLA Beats Baselines on All Nine Real-Robot Tactile Tasks

A robotics preprint out of NeoteAI and Fudan's TEAI group has a claim worth reading carefully, because the scanner-generated headline overstates what the paper actually reports. The authors describe N0-VTLA as the first vision-tactile-language-action foundation model pretrained on tactile data at scale, and say it beats the strongest baseline on every one of nine real-robot tasks in their NeoReal suite. That is 'wins every task' in the relative sense, not 'solves every task'. The mean real-world success rate is 47.2%, versus 29.4% for π0.5, the closest competitor.

Why the distinction matters: contact-rich manipulation, the kind where a gripper has to feel deformable objects or slot parts together by touch, is where vision-only stacks routinely fall over. The pitch is that adding tactile into the foundation-model recipe from the pretraining stage, rather than bolting it on downstream, lifts numbers across the board. On the twenty-task NeoSim benchmark N0-VTLA reports 63.8% average success against 44.0% for π0.5, and on UniVTAC it averages 83.1% versus 67.1% for InternVLA-A1. Those are meaningful gaps if they hold up outside the authors' own evaluation setup.

The team also introduces ALTER, an advantage-conditioned offline reinforcement learning method that converts trajectory comparisons into binary advantage labels and improves policies from stored deployment data. On three long-horizon real-robot tasks, ALTER-trained N0-VTLA policies reach 75-95% success. That is the more practically interesting result, because it points at a loop where deployed robots keep getting better from their own operational data rather than needing fresh demonstrations for every new task.

The honest caveat is that this is a single team's benchmark, on a dataset and task suite the same team built, without independent replication yet. What the paper does not give you here is total parameter count, compute cost, or evidence that the recipe transfers to tactile hardware other groups already use. Take the specifics as reported, not settled.

For anyone building dexterous manipulation systems, the signal is directional rather than definitive: tactile-scale pretraining is starting to earn its own place alongside vision and language in the foundation-model stack, and the teams that adopt it early will be the ones mapping the actual generalization limits.