Panasonic, Berkeley and Stanford unveil T-Rex robot AI that corrects grip from fingertip touch
The system pairs slow camera-and-language planning with faster fingertip corrections and posted a 65% success rate on 12 contact-rich tasks, Panasonic said.
Panasonic Holdings said Wednesday that it co-developed T-Rex, a robot control system that uses fingertip touch to fix its grip in real time, with researchers from UC Berkeley, Stanford, NVIDIA and other partners.
T-Rex stands for Tactile-Reactive Dexterous Manipulation. Most robot helpers decide moves from camera images and spoken or written instructions, a setup called a vision-language-action model. Those models often update only a few times per second, which is too slow when a soft object slips or a key sticks.
T-Rex keeps that slower planner at about five updates per second and adds a fingertip-touch corrector that runs about 20 times per second. In lab tests on 12 contact-heavy jobs — turning pages, moving an egg, unlocking a door, screwing in a bulb — Panasonic said T-Rex succeeded 65% of the time, versus 35% for the prior best baseline and 42% when touch was turned off.
The team trained on more than 7,700 two-handed robot episodes and will present the work at the Conference on Robot Learning in Austin next week. Company claims of success rates are Panasonic’s.
Source: Panasonic Holdings press release, Nov. 5, 2026; T-Rex project site; arXiv:2606.17055