The Collapse of Constrained Physical AI

Abstract

The most profound impact of AI will not occur on a screen, but within the physical systems that define our world. From autonomous flight and last-mile delivery to the surveying of active construction zones, systems that currently require constant human intervention are moving fast towards being autonomous controlled by large models. Yet, moving large-scale intelligence from the cloud to the edge introduces a unique set of challenges. This talk examines the constraints that large models must meet in order to be effectively deployed in physical environments. The talk then explores how enforcing these constraints can lead to a collapse in both task success and system safety. Finally, the talk presents recent work from a collaboration between the Edge Computing Lab and FieldAI that restores both safety and success in compressed Vision-Language-Action (VLA) models.

Bio

Jason Jabbour is a PhD student in Computer Science at Harvard University and an NSF Graduate Research Fellow. He is a member of the Edge Computing Lab, advised by Professor Vijay Janapa Reddi, and has previously conducted research at FieldAI. His research lies at the intersection of machine learning, computer systems, and robotics, with a focus on improving the safety and efficiency of physical AI systems. His work studies the emerging dependency between high-level physical AI capabilities and the efficiency of the underlying computing infrastructure. In particular, he investigates how the reasoning and contextual understanding of large-scale models can be leveraged for robotics, enabling solutions to previously intractable problems, but only when supported by rigorous resource-efficiency, high-throughput inference, and optimized communication frameworks.