Panasonic Holdings, partners develop T-Rex robot manipulation AI; 65% average success rate across 12 tasks

Panasonic Holdings and Panasonic R&D Company of America (PRDCA) announced on Nov. 5, 2026, that they had jointly developed T-Rex, a robot manipulation AI, with NVIDIA and universities. It achieved an average success rate of 65% across 12 real-world tasks involving touch, outperforming a comparison method at 35%. The rate was 42% when tactile information was excluded.

In Panasonic Holdings' evaluation, T-Rex achieved an average success rate of 65% when using tactile information, compared with 42% without it. Control that detects force and deformation at the fingertips is relevant to the challenge of assigning robots tasks that require adjustments based on contact, such as unlocking a door or retrieving a card. Panasonic Holdings says it is targeting component-handling tasks in manufacturing and handling flexible or fragile items in logistics, but the announcement does not report any deployments.

T-Rex’s focus is on combining, within the same robot manipulation system, a process that uses vision and language to determine the broad outline of a task and another that adjusts actions in response to changes in fingertip contact. RoboSta reported that the first runs at about 5 Hz and the second at about 20 Hz. Rather than repeatedly reprocessing camera footage, the design uses the tactile side to respond to slipping or excessive pressure during contact. This is consistent with the evaluation’s focus on tasks in which an object’s state changes after contact. However, 65% is the average across 12 tasks; it does not show results for individual tasks or success rates in real-world operation. The data composition also offers clues to the design’s significance. The system learns visual and motion knowledge from 22,889 hours of human egocentric video, learns contact actions from 100 hours of robot data with tactile information, and is then adapted using about 100 demonstrations for each task. This setup separates the acquisition of broad motion knowledge from learning robot-specific tactile control and adapting to individual tasks. The comparison showing that the average success rate fell to 42% without tactile information indicates that touch affected performance under at least the conditions of this evaluation, but it does not quantify the separate contributions of touch and high-speed control. A related article in this newspaper on factory automation, “NVIDIA introduces coordinated AI agent design to investigate production problems in factories” (Nov. 4, 2026), concerns a different layer of processing. As its headline indicates, that article focuses on investigating production problems, while T-Rex focuses on manipulating objects through physical contact. There is no basis for treating the two as the same system or as part of a coordinated plan, but they show that information processing and control of physical work need to be assessed separately when discussing factory automation. Manufacturing and logistics are applications Panasonic Holdings has identified as targets; the announcement does not report deployments. The next points to check are the specifications of the robots and tactile sensors used in the experiments, the number of trials and success rates for each task, response latency and reproducibility under real-world conditions. Although the scale of the data and the training process have been disclosed, these details cannot be confirmed from the announced materials, and the average success rate alone is not enough to determine the scope of potential use in mass-production facilities.

Information such as force and deformation from 10 fingertip tactile sensors is synchronized with camera footage, robot states and actions, and natural-language instructions. Training starts with 22,889 hours of human egocentric video, followed by tactile data to teach robot-specific contact actions, and ends with adaptation using about 100 demonstrations for each task. RoboSta reports that action generation based on vision and language runs at about 5 Hz, while action adjustment based on touch runs at about 20 Hz. Evaluation tasks include turning pages, transferring eggs, unlocking a door and retrieving a card.