Customizing Robot Behaviour through Interactive Learning
NCRN Tech Talk
Speaker: Nils Wilde
Abstract:
In many applications, robots are required to simultaneously optimize their behaviour for different competing goals. For instance, mobile platforms should complete delivery tasks as quickly as possible while navigating safely and respecting social norms. Thus, a central problem in human-robot interaction (HRI) is how users who are not experts in robotics can specify complex objectives for autonomous robots. Different users have different preferences for the best trade-off between objectives. We study human-in-the-loop learning frameworks that allow users to customize robot behaviour to their preferences through a sequence of simple interactions, in particular choice feedback where users choose between two presented options. To learn a user's preference within a few such iterations, the robot needs to actively select new options to be presented to the user. Therefore, we are interested in exploring different optimal trade-offs for multi-objective optimization problems, i.e., the set of Pareto-optimal solutions. The applications of our work include high-level motion planning in human-centered environments, manipulation in servicing tasks, environmental monitoring missions, as well as multi-robot pickup and delivery.
Bio:
Nils Wilde is currently a Postdoctoral Fellow in the Autonomous Multi-Robots Lab working with Javier Alonso-Mora at TU Delft. Until August 2021 he was a postdoctoral fellow at the Autonomous Systems Lab at the University of Waterloo where he also did his PhD in Electrical and Computer Engineering (ECE) under the co-supervision of Dana Kulić and Stephen L. Smith from 2016 to 2020. Before that he completed his BSc. and MSc. degrees at the Technical University Berlin in 2012 and 2016, respectively.
Nils' research combines robot motion planning and human robot interaction (HRI), investigating how inexperienced users can define complex behaviours for autonomous mobile robots via active learning frameworks. Recent work broadens the focus to high level coordination of multi-robot systems under uncertainty as well as theoretical work on multi-objective optimization for robot planning problems.
Ìý