Advancing physical AI: Building robots that learn from failure

I enjoy sports and have played tennis since the fifth grade through university. While I didn’t have a clear career path in mind, I hadn’t been able to achieve my goal of competing in a national tennis tournament in high school, so I joined the School of Informatics in university seeking an environment where I could continue to compete while balancing academics. My interest in AI began to grow through my undergraduate research. I wanted to combine AI, which was gaining increasing attention at the time, with my longtime experience in tennis, going on to pursue research under the theme of “AI for Tennis”. In tennis, players make strategic decisions based on their opponent’s position and the likely direction of the next shot. In my thesis, I used AI to explore whether this information could be quantified through motion capture and applied to strategic decision-making. Specifically, I collected and analyzed data such as players’ positions (where they stand and where they hit the ball), as well as the ball’s trajectory and bounce points, aiming to contribute to performance improvement in matches. Though I continued to struggle to choose a career even after entering graduate school, my research allowed me to discover the excitement of visualizing phenomena previously understood only intuitively. I also became fascinated by the potential of technology to expand human capabilities, which inspired me to contribute to the creation of new technologies as a researcher.

I am currently a part of the Physical AI laboratory, which was established in April 2026. The laboratory aims to realize a world where humans can safely coexist and collaborate with a wide range of robots across diverse real-world environments. It also promotes a vision of evolving robots to contribute to the development of society as a whole by incorporating on-site knowledge, advancing collaborative capabilities, and sharing these outcomes across locations and operations. To achieve this vision, we are developing Fujitsu Kozuchi Physical OS (*1), a platform that integrates robots, sensors, systems, and physical spaces to enable the robots to correctly understand tasks and execute the appropriate actions based on operational instructions.

Within this initiative, I leverage my experience in computer vision and video understanding to conduct research on robotic intelligence (*2). My current focus is on technologies that enable the reliable deployment of Vision-Language-Action (VLA) models (*3), which serve as the decision-making component of robotic systems. Key challenges include handling errors that arise during long-horizon tasks and adapting to unexpected changes in the environment. To address these challenges, I aim to develop technologies that enable robots to proactively avoid failures and, when they occur, accurately identify their causes and recover autonomously. I believe that overcoming these challenges will pave the way for technologies that can operate reliably in real-world environments.

Titles, numerical values, and proper nouns in this document are those reported when this interview was made.