Redesigning Perception and Computation for Energy-Efficient Physical AI

We aim to realize "AI that can understand the real world as it is and act instantly,” and to uncover the principles of intelligence that operates in the physical world. Current vision-based AI relies on frame-based cameras, treating inherently continuous time and motion as discrete snapshots, which limits the integration of prediction and action. We discard this assumption. Starting from event cameras that capture only moments of change, we build a spatiotemporal perception framework that directly handles the flow of time. Our goal is to enable physical AI that does not “sense, then think,” but rather “thinks while sensing and acts simultaneously.” Furthermore, by co-designing semiconductor hardware to match this new form of perception, we seek to develop biologically inspired, energy-efficient intelligent systems, and to open up further new imaging applications.