A camera can show a hand move around a test tube, but not how pressure changes when its fingers make contact. In an Oct. 7 arXiv paper, researchers report that training a robot with human video-and-touch data raised its average success rate from 22.5% to 57.5% across four real-world tasks.
The team’s TouchScale dataset contains 500 hours of human interactions, about 87,000 episodes and roughly 2,000 task descriptions. A head-mounted camera that records color and depth, two wrist cameras and tactile gloves on both hands share a synchronized timeline. Each glove measures pressure at 880 points across the fingers and palm.
The recordings have no labels for the actions people take, so the researchers did not train the model to copy human movements. Instead, they used the video, task instructions and current touch readings to teach a tactile-prediction part of the model to predict how contact would change over the next movement. The model's action module also uses that predicted-touch signal when it predicts robot movements. Both versions started from the same N0-VTLA checkpoint; one received this intermediate training and the other skipped it. Both were then trained on the same 50 teleoperated robot demonstrations per task.
On one xArm6 robot arm fitted with a BrainCo Revo 2 hand, the TouchScale-trained version succeeded in an average of 57.5% of trials across the four tasks, compared with 22.5% for the baseline without TouchScale mid-training, a difference of 35 percentage points. The tasks were sorting soft from hard objects, removing a bottle cap, transferring a test tube and wiping a whiteboard. Each task was tested in 20 trials.
The evidence is limited to one robot platform and four tasks. The dataset has no human action labels, and the authors say results across other robot designs remain untested. They also found only a weak relationship between predicting future touch and predicting actions, leaving the role of touch prediction in robot control uncertain. The paper is an arXiv preprint.
The full dataset is not yet available. Its Hugging Face page currently lists 100 of the 500 hours, with the complete release planned before November 2026. Users must agree to share contact information to access the files. The dataset is licensed CC BY-NC 4.0 for noncommercial research and education; the maintainers ask commercial users to contact them.
Sources
- Original TouchScale paper, posted Oct. 7, 2026
- Full paper and experimental results
- TouchScale dataset page, release status and access terms
- Image source and license record
Illustrative photo: XELA Robotics’ uSkin sensors fitted to a separate Allegro hand; this is not the TouchScale study hardware. Photo: XELA Robotics / Wikimedia Commons, CC BY-SA 4.0.