This project, developing new techniques for enabling planners to automatically learn from experience, offers fast, high-quality solutions to very large motion planning problems that arise in robotics, CAD/CAM, animation of virtual characters, and surgical planning. These problems are challenging because they require searching high-dimensional state spaces with complex geometric constraints, nonlinear dynamics, often with contact and impact, and long time horizons. Prior approaches have sought to solve these problems efficiently by embedding a great deal of domain knowledge into planning, but have relied heavily on human expertise to develop and exploit this knowledge. This process is tedious, error-prone, and does not scale well to harder problems that do not possess an obvious structure. This project will investigate automated strategies for planning systems to automatically discover common solution structures from past experience and to reuse this knowledge in new problems, with the ultimate goal of demonstrating a system that automatically optimizes planning strategies for a novel domain with minimal training and hand tuning.
Broader Impacts: The ability to solve large planning problems efficiently has myriad benefits to many fields of knowledge. But on a human level, this grant will provide the opportunity to recruit an undergraduate or Master's student intern from a minority-serving institution each summer, in cooperation with the Alliance for the Advancement of African-American Researchers in Computing (A4RC). Broader dissemination of the work will be achieved by distributing a research and educational software library for task-and-motion planning (PyTAMP), and integrating software with the open-source ROS and OMPL libraries. Research will be integrated with education in robotics and AI courses at the undergraduate and graduate level, as well as in K-12 robotics outreach.
