Sankei Shimbun reports Panasonic Holdings will strengthen physical AI research and development

Sankei Shimbun reported on Nov. 11, 2026, that Panasonic Holdings had announced plans to strengthen its research and development of physical AI.

Viewed alongside existing research results, the report highlights different challenges in enabling robots to understand the physical world in areas we have covered at Panasonic Holdings. Our Sept. 10, 2026, article, Panasonic Holdings unveils Proxy3D at CVPR 2026 to enhance VLMs’ 3D capabilities, covered 3D perception by VLMs. In contrast, our Nov. 5, 2026, article, Panasonic Holdings develops physical AI for robots using tactile sensing, with a 65% task success rate, focused on recognizing contact conditions and adjusting movements. The average success rate in the experiments was 65%, and the training data comprised more than 7,700 items and about 100 hours of data. Spatial perception and contact-based manipulation are related, but the former’s results do not necessarily lead directly to success in the latter’s tasks. The summaries of those previous articles do not confirm results from combining both in the same real-world task. What can be taken from this is that there remains a gap to examine between the breadth of research themes and demonstrations that link individual results into a sequence of actions in the field. 3D recognition is involved in identifying objects and their positions, while tactile information helps determine their state after contact. To assess practical utility, the focus will be not only on individual recognition and manipulation metrics but also on whether a complete task can be performed when objects or environments change. However, there is also no basis at this point to conclude that such integrated tests have not been conducted. The more than 7,700 items and about 100 hours of data used in tactile learning indicate the specific scale of the research inputs, but do not, on their own, demonstrate the scope of applicability or reproducibility. Nor can the average success rate of 65% be treated as field performance without taking into account the experimental conditions and types of tasks. Going forward, we will be watching how spatial recognition and tactile control are connected in specific processes, how evaluation conditions are set, and what metrics are used if the research results are transferred to products or services. This article does not reconstruct the specific descriptions or statements in Sankei Shimbun’s report; readers should consult the original source for details.