Abstract: Current AI systems can synthesize videos, pass the bar exam, and write code. 
Despite these advances, robots still struggle with basic physical tasks, like folding a shirt, that 
humans perform naturally. The main reason is the Robotic Data Gap: Robotics has no internet. 
While digital AI trains on billions of hours of web data, robot learning relies on small, costly 
datasets that are difficult to standardize and highly heterogeneous. In contrast, humans are 
remarkably data-efficient, generalizing effortlessly from limited experience. This raises a key 
research question: Can we bridge this gap by building Physical AI systems that perceive, 
reason, and adapt to the physical world, driving data efficiency and scalable generalization?
In this talk, I will present recent efforts in Physical AI to integrate physical inductive biases, 
allowing robots to generalize beyond their limited training data. I will highlight ongoing work that 
incorporates structured representations, such as motion particles, object geometries, 
symmetries, and affordances, into learning-based robotic models. My work, spanning from 
manipulation arms to humanoids, demonstrates that this structure is the key to unlocking 
Embodied AI despite the constraints of real-world data scarcity.

Bio: Roei Herzig is a Postdoctoral Scholar at UC Berkeley and a Research Scientist at the MITIBM Watson AI Lab. He is advised by Professor Trevor Darrell and works closely with 
Professors Jitendra Malik, Shankar Sastry, and Deva Ramanan. Roei earned his PhD from Tel 
Aviv University under the supervision of Professor Amir Globerson. His research focuses on 
embedding physical inductive biases and structured representations into learning models to 
drive data efficiency and scalable generalization in Physical AI. He has been recognized with 
several distinctions, including the 2023 Dissertation Award for the best AI thesis in Israel and 
the Israeli Excellence in Data Science Postdoctoral Fellowship