Talk Title: Action Hallucination in Generative Robot Policies. 

Abstract: Recent VLA models and diffusion-based robot policies are highly expressive, but robotics is not just another generative modeling problem. In this talk, I will discuss action hallucination: generated robot behaviors that look plausible to a model but violate physical feasibility or fail as executable/safe plans. For generalist robots, this is a foundational safety problem: a policy cannot be reliably safe if its action generator does not respect the structure of physical behavior. Drawing on our analysis of generative VLAs, I will argue that these failures often arise from structural mismatches between feasible robot behavior and common generative architectures, not merely from insufficient data. I will outline three barriers that help explain empirical failures in robot foundation models. I will then discuss implications for safer generalist robots, including structured action spaces and verification-guided test-time computation.

Talk Abstract: TBD

Talk Abstract: TBD