Learning to Search in Task and Motion Planning

Abstract: Task and motion planning (TAMP) problems in robotics are an instance of methods that combine logical reasoning and low-level motion planning, to enable complex, long-horizon behaviors. Despite significant progress in TAMP solvers over the last decade, several core problems remain open, one of them being that it is not clear what the role of learning should be to improve them. In this talk, we will show how to design TAMP solvers that learn to plan better and faster from past experience of solving similar problems in similar environments. We will show how a learning-guided TAMP system can plan long-horizon manipulation tasks in various environments, including in chemistry labs, where robot arms can execute a diverse range of multi-step chemistry experiments.

Bio: Florian Shkurti is an assistant professor in computer science at the University of Toronto, where he leads the Robot Vision and Learning lab (http://rvl.cs.toronto.edu/). He is a faculty member of the University of Toronto Robotics Institute, a Faculty Affiliate at Vector Institute for AI. His research group develops methods that enable robots to perceive, reason, plan, and act effectively and safely, particularly in dynamic environments and alongside humans. Application areas of his research include field robotics for environmental monitoring, visual navigation for autonomous vehicles, and mobile manipulation. He is the recipient of the Alexander Graham Bell Doctoral Award, the AAAI Robotics Fellowship, the Amazon Research Award in Robotics, and the Connaught New Researcher Award.