Robots that Learn and Plan — Unifying Robot Learning and Motion Planning for Generalized Task Execution

Robots have the potential to assist people with a variety of everyday tasks, but to achieve that potential robots require software capable of planning and executing motions in cluttered environments. To address this, over the past few decades, roboticists have developed numerous methods for planning motions to avoid obstacles with increasingly stronger guarantees, from probabilistic completeness to asymptotic optimality. Some of these methods have even considered the types of constraints that must be satisfied to perform useful tasks, but these constraints must generally be manually specified. In recent years, there has been a resurgence of methods for automatic learning of tasks from human-provided demonstrations. Unfortunately, these two fields, task learning and motion planning, have evolved largely separate from one another, and the learned models are often not usable by motion planners. In this thesis, we aim to bridge the gap between robot task learning and motion planning by employing a learned task model that can subsequently be leveraged by an asymptotically-optimal motion planner to autonomously execute the task. First, we show that application of a motion planner enables task performance while avoiding novel obstacles and extend this to dynamic environments by replanning at reactive rates. Second, we generalize the method to accommodate time-invariant model parameters, allowing more information to be gleaned from the demonstrations. Third, we describe a more principled approach to temporal registration for such learning methods that mirrors the ultimate integration with a motion planner and often reduces the number of demonstrations required. Finally, we extend this framework to the domain of mobile manipulation. We empirically evaluate each of these contributions on multiple household tasks using the Aldebaran Nao, Rethink Robotics Baxter, and Fetch mobile manipulator robots to show that these approaches improve task execution success rates and reduce the amount of human-provided information required.Doctor of Philosoph