Plan Projection, Execution, and Learning for Mobile Robot Control

Abstract

Most state-of-the-art hybrid control systems for mobile robots are decomposed into different layers. While the deliberation layer reasons about the actions required for the robot in order to achieve a given goal, the behavioral layer is designed to enable the robot to quickly react to unforeseen events. This decomposition guarantees a safe operation even in the presence of unforeseen and dynamic obstacles and enables the robot to cope with situations it was not explicitly programmed for.<br /> The layered design, however, also leaves us with the problem of plan execution. The problem of plan execution is the problem of arbitrating between the deliberation- and the behavioral layer. Abstract symbolic actions have to be translated into streams of local control commands. Simultaneously, execution failures have to be handled on an appropriate level of abstraction. It is now widely accepted that plan execution should form a third layer of a hybrid robot control system. The resulting layered architectures are called three-tiered architectures, or 3T architectures for short.<br /> Although many high level programming frameworks have been proposed to support the implementation of the intermediate layer, there is no generally accepted algorithmic basis for plan execution in three-tiered architectures. In this thesis, we propose to base plan execution on plan projection and learning and present a general framework for the self-supervised improvement of plan execution. This framework has been implemented in APPEAL, an Architecture for Plan Projection, Execution And Learning, which extends the well known RHINO control system by introducing an execution layer.<br /> This thesis contributes to the field of plan-based mobile robot control which investigates the interrelation between planning, reasoning, and learning techniques based on an explicit representation of the robot's intended course of action, a plan. In McDermott's terminology, a plan is that part of a robot control program, which the robot cannot only execute, but also reason about and manipulate. According to that broad view, a plan may serve many purposes in a robot control system like reasoning about future behavior, the revision of intended activities, or learning. In this thesis, plan-based control is applied to the self-supervised improvement of mobile robot plan execution.