· Research & technology· 日本語で読む
Panasonic Holdings develops physical AI for robots using tactile sensing, with 65% task success rate
Panasonic Holdings has developed physical AI that adjusts robot movements using tactile information. It recognizes contact conditions 20 times per second and achieved an average success rate of 65% in experiments. For training, the company collected more than 7,700 records of remotely operated video and tactile sensor data, totaling about 100 hours.
What happened
Panasonic Holdings announced on November 5, 2026, that it had developed physical AI that uses tactile sensing to adjust a robot's end-effector movements. The average success rate in experiments was 65%, which the company described as the world's highest performance.
Why it matters
Panasonic Holdings said it used tactile sensor information to adjust and train robot movements, achieving an average success rate of 65% in experiments. Its use of more than 7,700 records, totaling about 100 hours, that simultaneously captured video and tactile data underscores the need to examine data types and test conditions when evaluating end-effector tasks.
Analysis
The 65% success rate alone does not reveal which tasks were tested, how many trials were conducted or whether performance varied by task. At this stage, the assessment is limited to the development of a control method using tactile sensing and the reporting of an average success rate in experiments. The claim of "world's highest performance" should be attributed to Panasonic Holdings; without confirmation of the comparison group or evaluation conditions, it is not possible to conclude that the system is superior across a broad range of tasks.
The technical focus is the combination of a system that recognizes contact conditions 20 times per second with training that uses synchronized video and tactile recordings, in contrast with conventional AI that has often relied on camera images alone. The company disclosed a dataset of more than 7,700 records, totaling about 100 hours, but the breakdown by task and whether the system works with objects and environments not used in training cannot be confirmed from the information provided. The scale of data collection should be considered separately from reproducibility under unfamiliar conditions.
A related article in this newspaper, STMicroelectronics plans more than 20 demos at EdgeTech+ 2026 (November 5, 2026), says the company plans to feature embedded devices for robotics centered on STM32, integration with Isaac Sim and Sim-to-Real. While that article focuses on development environments and links with simulation, this announcement centers on control and training using tactile data that captures contact with physical objects. However, there is no information indicating that the two are technically connected. Another article, Robosta announces online seminar explaining NVIDIA's humanoid development platform (October 21, 2026), reports plans to cover GR00T, Cosmos, Isaac and Jetson Thor. Even as NVIDIA-related development platforms draw attention, there is no basis to conclude that this work was implemented on those platforms.
Three points warrant follow-up. First, the types of tasks, number of trials and comparison conditions used to calculate the 65% rate. Second, performance on tasks beyond cutting paper and squeezing a sponge, and with different objects and environments. Third, the task-by-task breakdown of the training data and how performance changes as more data is added. Without this information, it is difficult to assess whether the experimental results translate into generality or reproducibility in real-world use.
Sources
Translated from the Japanese original: /news/panasonikku-hd-shokkaku-wo-tsukau-fijikaru-ai-wo-kaihatsu-to-0c65b9. The Japanese version is authoritative if they differ.