News · 2026-08-03

Robot Policies That Predict the Touch Before They Make It

Two technical reports released together by NeoteAI and Fudan University give robot manipulation policies a form of touch that predicts contact instead of merely reporting it. The first, N0-VTLA, wins all nine real-robot tasks in the authors' benchmark and reaches 63.8 percent mean success on a twenty-task simulation suite against 44.0 percent for the strongest baseline. Both releases ship code and checkpoints.

Key facts

• N0-VTLA reports 63.8 percent mean success across twenty simulated tasks, against 44.0 percent for the strongest baseline, and wins all nine real-robot tasks.
• Policies trained with the authors' offline improvement method, ALTER, reach 75 to 95 percent success on three long-horizon real-robot tasks.
• The shared dataset covers more than 30,000 hours of vision-tactile data across six robot bodies and 450 tasks, with a 5,000-hour open subset.
• Primary sources: N0-VTLA, arXiv:2607.23782 and N0-TWAM, arXiv:2607.23783, both dated 25 July 2026.

The authors state the pretraining claim directly: "to our knowledge, N0-VTLA is the first VTLA model pretrained on tactile data at scale." That phrase, at scale, is what separates this from a decade of tactile sensing demos.

How it works

The interesting design decision in N0-VTLA is what it refuses to feed the policy. Raw tactile images never enter the model's vision-language context. Instead the system takes the current fingertip gel-sensor image, subtracts a zero-contact reference image of the same fingertip, and encodes what is left, which is the contact itself rather than the sensor's idiosyncrasies. A small predictor then combines those contact tokens with vision and language to estimate the net tactile change over the next chunk of motion. Only that prediction reaches the action expert.

The distinction is the whole idea. Current touch is a report from the past: we have hit something. The predicted latent is a forecast: this is the contact the next few centimetres of motion will produce. A policy that only knows the first can react. A policy that has the second can avoid.

The behaviour it buys is easy to picture. When a plug catches on the rim of a socket, the tactile policy lifts, realigns, and retries, instead of continuing the downward push that vision planned. On a squeezable bottle it makes small aperture adjustments rather than crushing it. Anyone who has watched a vision-language-action model confidently push a part into the wrong place will recognise the failure being fixed.

The second bet

N0-TWAM approaches the same problem from the world model side. Rather than predicting only what the camera will see next, it jointly generates a future camera stream, a future tactile stream, and the action that produces both. That forces its internal model to represent whether a grasp is stable, whether a plug is seated, whether contact is sliding into a slip. A model that only predicts pixels can generate a plausible video of a successful grasp; a model that must also predict the force pattern cannot fake it as easily.

It keeps two tactile paths, and the split is clean: predicted touch prevents a bad contact, observed touch catches and corrects one that is already happening. Removing either path measurably hurts the authors' results, which is evidence both are doing work.

Its other idea is what the authors call tactile punctuation. Contact onset, secure grasp, release, seating, and loss of load carve a long demonstration into subtasks. At runtime a predicted event can propose moving on, but real observed touch has to confirm it. A gripper closing on empty air therefore does not get mistaken for a successful grasp, which is one of the most common ways long-horizon manipulation quietly derails.

Why it matters

The claim worth taking from this is not that robots can now feel. It is that the last centimetre of manipulation needs a model of contact dynamics, and more pixels will not supply one. Vision can see a plug near a socket; touch can tell whether it is jammed. Vision can see a gripper close; touch can tell whether it loaded anything. Vision can plan a wipe; touch can regulate force against the surface for the whole stroke.

That reframes what tactile data is for. It is not an extra sensor channel to concatenate. It is the only signal that carries the information the task actually turns on, which is why pretraining on 30,000 hours of it matters more than any single benchmark row. Compare with Gemini Robotics 2 controlling a humanoid from feet to fingertips and with HiFi-UMI dropping the robot from robot training data, both of which are attacking the same last-centimetre problem from other directions.

The honest caveat: every number here is author-reported, from author-designed benchmarks, with no independent replication and no peer review. The real-robot evaluations run a small number of trials per task, and the papers themselves warn against reading individual task bars too closely. The releases are stronger than demo-only announcements because checkpoints, code, and deployment paths are all published, and the project pages include the material to check the claims. Somebody outside the lab still has to do it.

Primary source, verified: read the paper → (arXiv 2607.23782)

Ground Truth. (2026, August 3). Robot Policies That Predict the Touch Before They Make It. Ground Truth. https://groundtruth.day/news/robot-policies-that-predict-the-touch-before-they-make-it.html

@misc{groundtruth:robot-policies-that-predict-the-touch-before-they-make-it,
  title  = {Robot Policies That Predict the Touch Before They Make It},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/robot-policies-that-predict-the-touch-before-they-make-it.html}
}

Topics: robotics · tactile-sensing · vision-language-action · world-models · manipulation · open-source