Joint Active Learning of Feasibility Classifiers and Skill Policies for Robot Task and Motion Planning

Abstract

Recently, viral demonstrations from robotic companies have indicated that robots can perform intricate actions and solve complex tasks. However, many of these demonstrations are hand-designed and the robots are essentially executing a fixed sequence of actions. Therefore, the main barrier preventing the widespread adoption of robots is their inability to reason and develop an action plan given a task. Task and Motion Planning (TAMP) is a promising method to perform robotic decision-making in long-horizon, complex tasks as it introduces abstractions over the states and actions that simplify the planning problem. However, a core issue in TAMP approaches is their inability to determine if a robotic plan is feasible to execute without actually running it in the real-world. Some prior works attempt to learn classifiers that determine if a robotic plan is feasible or not, but are limited due to either requiring offline training or assuming the abstract actions are well trained. The main contribution of this thesis is designing a Bilevel Active Lifelong LEarning Robot, or BALLeR, that jointly learns a robotic plan feasibility classifier and improves the quality of the abstract actions for solving a particular task. Additionally, BALLeR allows the robot to autonomously gather information and learn. We evaluate BALLeR across several different simulated benchmarks and provide extensive analysis on its performance and the relationship between the abstract level classifier and low level policy.