Multi-Objective Motion Planning and its Application to Active Preference Learning

Abstract: Robot motion can be evaluated according to multiple objectives, such as path length, speed, legibility, and distance from obstacles.  When designing a motion planner, one must decide how to trade-off these different objectives.  To complicate matters, robots are often operating amongst humans, whose preferences on robot behaviour may differ.  These preferences can be thought of as a weighting on the different objectives.  In this talk, we discuss approaches for solving multi-objective motion planning problems and exploring trade-offs between objectives.

Linear scalarization is a common approach for solving such multi-objective optimization problems.  This approach involves formulating a single cost function made up of the weighted sum of the individual objectives. However, determining the influence of different weight vectors on system behaviour can be complex.   We discuss how to compute a finite set of weight vectors that offer a comprehensive representation of potential system behaviours.  A limitation of linear scalarization is its lack of Pareto-completeness;  there are Pareto-optimal solutions that are not the solution of the scalarized optimization problem for any vector of weights.  To overcome this, we study an alternate form of scalarization based on a weighted maximum of objectives.  We demonstrate the application of this scalarization in robot motion planning and the learning of user preferences.

Bio: Stephen L. Smith is a Professor in Electrical and Computer Engineering at the University of Waterloo, Canada, where he holds a Canada Research Chair in Autonomous Systems. He is a the Co-Director of the Waterloo Artificial Intelligence Institute (Waterloo.AI) and the Director of the Autonomous Systems Lab. Prior to Waterloo, he was a postdoc in the Computer Science and Artificial Intelligence Lab (CSAIL) at MIT. He received his BSc degree from Queen’s University, his MASc degree from the University of Toronto, and his PhD degree from the University of California, Santa Barbara. Prof. Smith is a Professional Engineer and serves as an advisor for several startups in transportation and robotics including RideCo and Swap Robotics. He has served on the editorial board of the IEEE Transactions of Robotics, the IEEE Transactions on Control of Network Systems, and the IEEE Open Journal of Systems and Control. He was a member of the organizing committee for the 2024 American Control Conference, the 2021 IEEE RO-MAN and the 2026 MTNS. Prof. Smith has received several awards including the Early Researcher Award from the Province of Ontario, the NSERC Discovery Accelerator Supplement Award, and three Outstanding Performance Awards from the University of Waterloo. His main research interests lie in control and optimization for autonomous systems, with a focus on safe motion planning, future mobility-on-demand systems, and the interaction of humans with autonomy.