Panasonic's T-Rex robot AI feels with its fingertips to screw in bulbs and turn pages

People pull out a thin card or turn a key by feeling force, slip and give through their fingers, not just by looking. Vision-language-action (VLA) models, which turn camera images and instructions into robot motion, mostly rely on cameras, and when they do use touch they tend to treat it as a static input[1]. Panasonic also points to two practical gaps: there is little large-scale training data that includes touch, and vision-language models update too slowly for quick tactile reactions[1].

T-Rex splits the job. The vision-and-language part plans the overall motion at about 5 Hz, while a tactile module reads contact at about 20 Hz and makes fine corrections, so the robot can respond to slipping or pressing too hard without rerunning slow image recognition[1].

Training happened in three stages. The model first learned general visual and motion knowledge from 22,889 hours of first-person human video, then learned contact-heavy robot behaviour from more than 7,700 episodes, about 100 hours, of two-handed manipulation covering over 200 everyday objects and 22 basic actions[1]. Each episode records head and wrist camera video, the robot's states and actions, force and deformation data from tactile sensors on 10 fingertips, and a natural-language instruction. Finally, it was fine-tuned on roughly 100 demonstrations per task[1].

On 12 real-world tasks, including turning pages, transferring eggs, applying toothpaste, unlocking with a key, pulling out a card and screwing in a light bulb, T-Rex averaged a 65% success rate, against 35% for the best previous method[1]. With touch input removed it dropped to 42%, which the team says confirms that the tactile sensing and fast reactions are doing real work[1]. Videos of the tasks are on the project site[3].

Panasonic wants to use the approach for jobs that need fine force control, such as inserting, removing and screwing components in manufacturing and handling soft or fragile items in logistics, to ease labour shortages on the front line[1]. A 65% average is still far from factory-ready, but almost doubling the previous best on contact-heavy tasks shows how much robot hands gain from feeling what they grip.