Panasonic HD’s T-Rex Tactile AI Hits 65% Success on Dexterous Two-Handed Robot Tasks
Panasonic Holdings said on November 5 that it has co-developed T-Rex, a robot manipulation AI that adjusts two-handed motions from fingertip touch almost instantly, according to the company’s global newsroom and Nikkei. Working with Panasonic R&D Company of America and researchers at UC Berkeley, NVIDIA, Stanford, Sapienza University of Rome and ItalAI, the team reports a 65% average success rate across 12 contact-rich real-world tasks, 30 percentage points above the previous best baseline.
T-Rex pairs a vision-language-action planner that updates at about 5 Hz with a tactile correction loop running near 20 Hz. Conventional models lean on camera frames and treat touch as a static cue, which fails when a card slips or a soft object deforms. The group collected more than 7,700 bimanual episodes—about 100 hours—covering 200-plus everyday objects and 22 basic actions, with head and wrist video, ten fingertip force and deformation channels, and natural-language instructions recorded together.
After mid-training on that tactile set and fine-tuning with roughly 100 demonstrations per task, robots handled chores such as turning pages, transferring eggs, applying toothpaste, unlocking a key, pulling a card and screwing in a light bulb. Without tactile input the average success rate fell to 42%, underscoring how much the high-speed touch loop matters for factory and logistics work that still depends on human hands.
What it means in practice
The result lands as Japan debates how to govern machines that act in the physical world, including MIC’s rewrite of AI business guidelines for agents and physical AI and industrial tie-ups such as Hitachi and Fanuc’s physical-AI partnership for factory floors. Panasonic says it wants to use the approach for inserting, removing and screwing parts on production lines and for handling fragile goods in warehouses, easing labour shortages without claiming the robot replaces human judgment.
A paper on T-Rex is scheduled for the Conference on Robot Learning in Austin from November 9 to 12. What to watch is whether Panasonic folds the stack into group manufacturing robots and how the 65% success rate holds up outside lab props.