Panasonic HD Co-Develops a Robot Manipulation AI That Performs Dexterous Bimanual Tasks by Reacting Instantly to Fingertip Touch

Osaka, Japan, November 5, 2026 – Panasonic Holdings Corporation (Panasonic HD) and Panasonic R&D Company of America (PRDCA), together with researchers from the University of California, Berkeley, NVIDIA, Stanford University, Sapienza University of Rome, and ItalAI, have developed "Tactile-Reactive Dexterous Manipulation (T-Rex)," a robot manipulation AI that performs dexterous bimanual tasks while responding instantly to fingertip touch. T-Rex integrates fingertip tactile information into a VLA (Vision-Language-Action) model ⁽*¹⁾, which generates robot actions from visual and language information, and corrects its actions in real time as contact conditions change. In real-world evaluations across 12 tasks requiring delicate force control and the handling of soft objects, T-Rex achieved an average success rate of 65%, outperforming existing methods by 30 percentage points and achieving state-of-the-art performance ⁽*²⁾. These results will be presented at the Conference on Robot Learning (CoRL), an international conference in the field of robot learning, to be held in Austin, United States, from November 9 to 12, 2026.

When performing tasks such as pulling out a thin card or inserting and turning a key, people instantly adjust their movements based not only on vision but also on the force, slippage, and deformation they feel through their fingertips. Conventional VLA models, however, rely mainly on camera images, and even when they use tactile information, they often treat it as static input. This has made it difficult for them to handle tasks that require fine force adjustment or responses to changing contact conditions. Further challenges include the scarcity of large-scale training data containing tactile information, and the gap between the low update rate of vision-language models and the high control frequency needed for fast tactile responses.

T-Rex integrates low-frequency "action generation based on vision and language" with "high-speed action correction" driven by tactile feedback from the fingertips into a single model. While the motion generation module, which processes vision and language, generates the overall motion at approximately 5 Hz ⁽*³⁾, the action correction module, which processes tactile input, captures contact conditions in real time at approximately 20 Hz and finely corrects the action. This allows the robot to respond quickly to slippage or excessive pressing during contact without having to rerun high-latency image recognition each time.
For training data, the research team combined more than 200 types of everyday objects with 22 types of basic actions to collect over 7,700 episodes—a total of 100 hours—of bimanual robot manipulation data. Each episode synchronously records video from head and wrist cameras, the robot's states and actions, force and deformation information obtained through tactile sensors on 10 fingertips, and natural-language instructions. T-Rex first learns broad visual and motion knowledge from 22,889 hours of human egocentric video, then learns robot-specific contact-rich behaviors from this tactile-annotated data (mid-training), and is finally fine-tuned with approximately 100 demonstrations per task. In evaluations on 12 contact-rich 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 achieved an average success rate of 65%, far exceeding the 35% achieved by the previously best-performing baseline. Without tactile information, the average success rate dropped to just 42%, confirming the effectiveness of tactile sensing and the high-speed reactive mechanism.

Going forward, we aim to leverage this technology to enable robots to assist with tasks that require fine force control and adjustments based on contact conditions—such as inserting, removing, and screwing components in manufacturing, and handling soft or fragile items in logistics—thereby reducing the burden on frontline workers and helping to alleviate labor shortages.

Title: T-Rex: Tactile-Reactive Dexterous Manipulation
arXiv: https://arxiv.org/abs/2606.17055
Project Website: https://tactile-reactive-dexterous.github.io/
This research is a collaboration among the University of California, Berkeley, NVIDIA, Stanford University, PRDCA, Sapienza University of Rome, and ItalAI.

Notes
*1 An AI model that understands images and natural-language instructions and generates robot motions.
*2 Based on evaluations across 12 real-world tasks involving physical contact, including page turning, transferring an egg, applying toothpaste, unlocking a door, removing a card, and screwing in a light bulb. (PanasonicHD research, as of November 5, 2026)
*3 Hz is a unit indicating the number of processing or update cycles per second. 5 Hz means 5 updates per second, and 20 Hz means 20 updates per second.

• Panasonic×AI website
  https://tech-ai.panasonic.com/en/
• Panasonic×AI on X
  https://x.com/panasonic_ai