Radar / AI robotics / T-Rex
T-Rex: a robot manipulation AI that reacts to fingertip touch
Panasonic, UC Berkeley, NVIDIA and Stanford built T-Rex (Tactile-Reactive Dexterous Manipulation), a VLA model that corrects bimanual robot actions in real time from fingertip touch, hitting 65% average success across 12 delicate tasks versus 35% for the best baseline.
Why it matters
First VLA architecture to give touch its own high-frequency lane: a slow action expert plans at ~5 Hz while a tactile expert corrects the grip at ~20 Hz, beating the strongest baseline by 30 points on tasks like transferring eggs and screwing in light bulbs.
What you could build with it
A robotics startup prototyping warehouse picking could use the open 100-hour T-Rex dataset as mid-training data for its own tactile policy, teaching grippers to handle fragile items like eggs and glassware without crushing them.
Does it hold up?
Too early to judge for production use: a research system evaluated on one bimanual platform, with the 100-hour tactile dataset open-sourced; no independent team has deployed it outside the lab yet.
Built with T-Rex
- Panasonic HD Co-Develops a Robot Manipulation AI That Performs Dexterous Bimanual Tasks by Reacting Instantly to Fingertip Touchpanasonic · Official announcement: 65% average success across 12 tasks, open-sourced 100-hour tactile dataset, to be presented at CoRL 2026 in Austin, Nov 9-12.
- Panasonic's T-Rex Robot AI Reacts to Fingertip Touchinnovationessence · Writeup of the Nov 5 announcement, the Berkeley/NVIDIA/Stanford collaboration, and the CoRL presentation dates.
- Panasonic T-Rex Robot AI Uses Touch to Improve Manipulationbitcoinversus · Coverage noting the 5 Hz planning / 20 Hz tactile correction split, 7,700+ episodes over 100 hours, and the drop to 42% when touch is removed.
Learn more
First spotted on article: source.
More AI robotics
RoboParty RP1: a full-stack open-source humanoid robot for embodied AI researchRoboParty's 1.3-metre humanoid, launched 16 October and back in coverage on 2 November, comes in 24-DoF standard and…robotics · JEV 0.62Bear Robotics KINO-2: demonstration-to-skill robot-learning platformBear Robotics launched KINO-2 on Oct 21 at GTC Berlin: a robot-learning platform that turns human task demonstrations…robotics · JEV 0.59FineART: Scale AI and Hugging Face's open robotics dataset and plan-before-act VLA modelFineART is the largest open robot dataset with subtask labels: 40,543 episodes, 1,718 hours of footage, 151 tasks and…robotics · JEV 0.59Dream-1: robot foundation model that controls unseen bimanual robots zero-shot, at $3/hourDream Labs came out of stealth on 19 October with Dream-1, a robot foundation model offered through a public Dream API…robotics · JEV 0.52LeRobot v0.7.0: text steering for running robots, disk-less datasets, remote policy serverHugging Face shipped LeRobot v0.7.0, adding text steering of running robots, disk-less dataset streaming, a remote…robotics · JEV 0.48AGIBOT Innovation Day: GO-3 Preview, WITA-Omni, G2 Max, Link-U OS 1.0 and the Worken productivity metricShanghai embodied-AI firm AGIBOT unveiled its GO-3 Preview embodied foundation model, WITA-Omni interaction model…robotics · JEV 0.43
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.