Abstract: One of the most difficult challenges in robot motion planning is to account for the behavior of other moving agents, such as humans. Commonly, practitioners employ predictive models to reason about how other agents will move. Though there has been much recent work in building predictive models, no model is ever perfect: an agent can always move unexpectedly and in a way that is not predicted. In such cases, the robot may plan trajectories that appear safe but in fact lead to collision. Rather than trust a model's predictions blindly, we propose that the robot should also reason about the model's current predictive accuracy to inform the confidence in its predictions. This model confidence inference allows us to generate motion predictions that exploit modeled structure when it can successfully explain human motion, and degrade gracefully whenever the human moves unexpectedly. In this talk, I will present my recent work on:(1) how maintaining a Bayesian belief over a human model parameter can enable us to adapt conservativeness of predictions based on measurements, and(2) a general reachability-based framework that can reason about the potential future changes in the Bayesian belief itself.Finally, I will demonstrate some of the results from this work in the aerial and mobile robotics testbeds at Berkeley.

Bio: Andrea Bajcsy received her BS in Computer Science from the University of Maryland, College Park in 2016, and is currently a PhD candidate at the University of California, Berkeley in the Electrical Engineering and Computer Science Department under the direction of Professors Claire Tomlin and Anca Dragan. Her research focuses on combining control theory and machine learning to provide online safety assurances for robots operating around humans. At Berkeley, she collaborates with people in the InterACT Lab and the Hybrid Systems Lab and is supported by the NSF Graduate Research Fellowship.