Speaker: Max Simchowitz (Carnegie Mellon University)

Title: A Mathematical Basis for Moravec’s Paradox, and Some Open Problems

Abstract: Moravec’s Paradox observes that AI systems have struggled far 
more with learning physical action than symbolic reasoning. 
Yet just recently, there has been a tremendous increase in the 
capability of AI-driven robotic systems, reminiscent of the 
early acceleration in language modeling capabilities a few 
years prior. Using the lens of control-theoretic stability, this 
talk will demonstrate an exponential separation between 
natural regimes for learning in the physical world and in 
discrete/symbolic settings, thereby providing a mathematical 
basis for Moravec’s famous observation. We then explain the 
recent progress in robot learning by establishing that the 
innovations that immediately preceded these advances—
prediction of open-loop “action-chunks”, and use of generative 
models, such as diffusion models, to parametrize the 
conditional distribution of robot actions—directly mitigate the 
aforementioned difficulties. While our understanding of action 
chunking is rigorous, our findings regarding generative 
modeling are mainly empirical in nature. Thus, we conclude 
with open questions regarding how, under what conditions, 
and by what mechanisms popular generative models enjoy the 
properties uncovered in our experimental study.

Biography: Max Simchowitz is an assistant professor at the Machine Learning 
Department at Carnegie Mellon University with a courtesy appointment 
in the Robotics Institute. His work studies theoretical foundations and 
new methodologies for machine learning problems with an interactive, 
sequential, or dynamical component, currently focusing on 
reinforcement learning and applications to robotics. His past work has 
ranged broadly across control, theoretical reinforcement learning, 
optimization and algorithmic fairness. He received his PhD from 
University of California, Berkeley in 2021 under Ben Recht and Michael I. 
Jordan, and completed his postdoctoral research under Russ Tedrake in 
the Robot Locomotion Group at MIT. His work has been recognized with 
an ICML 2018 Best Paper Award, ICML 2022 Outstanding Paper Award, 
and RSS 2023 and ICRA 2024 Best Paper Finalist designations.