Short Research Statement

A hallmark of human intelligence is its remarkable data efficiency. A child learns to manipulate new objects after interacting with just a few toys, effortlessly generalizing their skills to new situations. In contrast, despite major advances in data-rich domains such as vision and language, modern AI has yet to produce comparably capable agents. This disparity stems from a core limitation: general-purpose agents lack the data comparable in scale and richness to the internet-sized datasets that catalyzed recent AI breakthroughs.

My research bridges this gap by building Physical AI systems that perceive, reason, and adapt to the physical world, driving both data efficiency and scalable generalization. To achieve this, I am pioneering Structured Physical Intelligence: designing architectures and learning algorithms grounded in physical inductive biases and structured world representations, using robots as the ultimate testbed.