CompCog: Physical Reasoning in Humans and Machines

While formal physics education typically begins in high school or college, humans develop sophisticated intuitions about the physical world long before entering a classroom. Even young children can predict whether a tower of blocks will fall over, how much weight a tree branch can support, or how far a ball will roll if kicked. This intuition about the physical world is a core part of human intelligence, contributing to our everyday commonsense knowledge, but it is still quite difficult to engineer systems that match the robustness and performance of this kind of human intelligence and that would enable robots and other machines to safely and intelligently interact with the real physical world. The current project aims to “reverse engineer” this aspect of human intelligence to determine what internal processes people are using to reason about the physical world. This project relies on a custom physics simulator that can express a wide range of possible physics, many of which differ from those experienced on Earth (e.g., gravity that is a little too strong or an unnatural relationship between force and motion). Human participants and AI machines will judge which physical laws seem most natural or correct and infer unseen parts of a scene using only the motion of visible elements. The comparison of human and machine strategies on identical tasks will uncover the representational commitments of each while advancing state-of-the-art methods for evaluating AI systems and providing inspiration for design of next-generation AI systems. The interdisciplinary nature of the project fosters workforce development at the cognitive science-AI frontier. Open-source tools, including physics engines, data analysis protocols, and datasets, will be freely accessible to the broader research community. This project advances knowledge about human intelligence and addresses NSF strategic priorities in AI research and understanding human intelligence and the White House’s “AI Action Plan” emphasis on preserving American leadership in transformative AI and maintaining global competitiveness in scientific domains of artificial intelligence, cognitive science, and education. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2545541 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Todd Gureckis | Institution: New York University, NEW YORK, NY | Award Amount: $913,250 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2545541 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2545541.html