Current Institution: University of California, Berkeley
Bio: Marcell Vazquez-Chanlatte is a final year PhD Candidate in Computer Science advised by Prof. Sanjit Seshia at the University of California, Berkeley. His dissertation work on inferring and teaching formal languages and programs using expert demonstrations lies at the intersection of Human-Robot Interaction, Grammatical Inference, AI-Safety, and Formal Methods. Prior to Berkeley, Marcell received a B.S. in Physics and a B.S. in Computer Science from the University of Illinois. Additionally, he has had software engineering and/or research internships at Google, Mozilla, Qualcomm, Toyota, and SRI International. Marcell’s current non-research interests include visual design, gardening, hiking, and being a dad.
Abstract: Specifications from Demonstrations: Learning, Teaching, and Control
Safety critical human-robot interactions are becoming increasingly common, with applications ranging from driving to manufacturing. This increased ubiquity brings an increased appetite for “auditable” systems whose high-level goals and assumptions can be easily communicated and analyzed by other agents, human collaborators, and regulators. My dissertation work considers the problem of learning and teaching formal task specifications using demonstrations. We will begin with an overview of the advantages of representing goals using logic/automata, e.g., compositional reasoning and decades worth of tool development for machine verification. Then, I will show that despite the intractably large (and discrete) nature of the hypothesis space, tasks modeled as task specifications, can be reliably and efficiently inferred from unlabeled and incomplete demonstrations. The resulting framework enables: (i) systems capable of safely refining tasks using user demonstrations and (ii) the synthesis of pedagogic examples for understanding specifications, e.g. task debugging and/or designing user on-boarding tutorials.