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 Title:  Safety in Physical Care: Context-Aware Embodied Intelligence for Caregiving Robots

Abstract: Physical caregiving tasks challenge conventional notions of robot safety: contact is often necessary, and safe behavior depends on the user, task, environment, and robot embodiment. In this talk, I will present research from the EmPRISE Lab on embodied intelligence for physical contact with humans. I will highlight our work on multimodal contact representations, personalized contact-aware control, and user-context-aware assistance, along with lessons from deployments with end users. Together, these efforts offer insights into the representations, learning algorithms, and safety principles needed for real-world physical caregiving robots.

Co-founder and CTO, Valgo

Talk Title:  Continuous Monitoring of Safety Performance 

Abstract: This talk focuses on safety decisions made during deployment of new robotic systems. I will begin by discussing what it means to be safer than humans in the context of autonomous vehicles, highlighting that this condition is necessary but not sufficient for deployment. To check whether AVs are safer than humans when deployed, it is natural to choose to continuously monitor their performance. However, this continuous monitoring can lead to some common statistical traps. I will conclude by discussing what these traps are and ways to avoid them.

Talk Title: Reliable and Scalable Evaluation for the Next Generation of Robot Policies

Abstract: Rapid progress in robot learning, foundation models, and large-scale datasets has given rise to robot policies that generalize across diverse tasks and environments. However, rigorously evaluating these policies remains a fundamental bottleneck. In this talk, I will present reliable and scalable methods for evaluating robot manipulation policies using imperfect simulators, including both physics-based simulators as well as video world models. Our approach builds on principled statistical techniques such as prediction-powered inference to combine a small number of real-world trials with large-scale simulation, yielding finite-sample, reliable guarantees on real-world performance. I will conclude with a brief discussion of open research questions on closing the loop between evaluation and policy learning to enable safe and robust robot policies.