CREO Seminar Talk: Yaqi Xie - Knowledge Guided AI Partner

Abstract

Building trustworthy AI partners alongside humans faces a reality that scaling alone cannot fix: embodied data is orders of magnitude scarcer than language and vision data. Before sufficient data can be collected, agents need to make smart use of what we have by reasoning with structure. This talk will show how structured knowledge and inductive bias can reduce reliance on brute-force data, enabling AI agents to understand context, generalize robustly, and collaborate with humans in the real world. The talk will introduce how structured knowledge can be embedded directly into neural architectures, integrating logical structure into representations rather than attaching reasoning as a post-hoc module. It will then demonstrate how knowledge-guided models support reliable contextual perception, including robust scene understanding and hallucination mitigation. Finally, it will present how grounded representations enable proactive human-robot collaboration across various real-world scenarios, where robots interpret human intent, anticipate actions, and coordinate behavior in dynamic environments. Together, these efforts outline a pathway toward capable, efficient, and trustworthy AI agents as genuine partners for humans, both digitally and physically.