Panasonic, UC Berkeley, Nvidia and Stanford researchers unveil T-Rex, a robot model that reacts to fingertip touch
The vision-language-action model plans at about 5 Hz and corrects with touch at about 20 Hz; Panasonic says it beat other methods on 12 real contact tasks, in its own evaluation.
Why it matters
Most vision-language-action models steer robots mainly from camera images and treat touch as static information, which makes fine force control hard. T-Rex splits the work: a slower module plans from vision and language while a faster one reads fingertip sensors and adjusts grip in real time, so slipping or pressing too hard can be caught without re-running image recognition. The tasks included turning pages, moving eggs, inserting and turning keys and screwing in light bulbs. Panasonic aims the method at factory insertion and screw-fastening and at handling fragile goods in logistics; watch for independent replication of the arXiv results.
Key facts
- Vision-language planning at about 5 Hz
- Tactile correction at about 20 Hz
- Training data on more than 200 everyday objects
- Evaluated on 12 real-world contact tasks
- Partners: UC Berkeley, Nvidia, Stanford, Sapienza University of Rome, ItalAI