Combining Navigational Planning and Reactive Control

Traditional AI methods for navigational planning use qualitative spatial representations and reasoning. Traditional robotics techniques for this task are based on numerical representations and reasoning. Recent work on robotics posits mechanisms for reactive control that directly map perceptions of the world to actions on it. This in turn has given rise to hybrid robot architectures that combine navigational planning and reactive control. But following traditional robotics techniques, navigational planning in these hybrid architectures too uses numerical methods. This raises the following question: Given a hybrid robot architecture, are numerical methods really needed for navigational planning? To explore this issue, we integrated a multistrategy qualitative navigational planner with a reactive-control mechanism. Then we embodied the integrated system on a physical robot. Next we gave the robot a series of navigation tasks in a visually structured spatial world containing discrete pathways, and monitored its actions as it executed the tasks in the presence of both static and moving obstacles. Our experiments show that for hybrid robots qualitative methods are su cient for navigational planning in at least one class of spatial worlds. Background, Motivations and Goals Spatial navigation is a classical problem in AI and robotics. Traditional AI methods for spatial navigation rely on deliberative planning based on qualitative spatial representations and reasoning [Davis 1986]. For example, STRIPS [Fikes and Nilsson 1971; Fikes, Hart and Nilsson 1972] combined the methods of meansends analysis and theorem proving to form qualitative plans. Its spatial representations captured topological relationships between spatial regions (e.g., rooms) This work has bene ted from many discussions with members of the AI and Robotics groups at Georgia Institute of Technology. We are especially grateful to Ron Arkin for allowing us to play with AuRA and Stimpy. This research has been partially supported by a CER infrastructure grant from the National Science Foundation (CCR-86-19886), and research grants from the National Science Foundation (IRI-92-10925) and the O ce of Naval Research (N00014-92-J-1234). and connections between them (e.g., doors) but did not specify any numerical measures such as distances. Recent AI programs for navigational planning (e.g., [Alterman 1988, Kuipers and Levitt 1988; Lawton and Levitt 1990; McDermott and Davis 1984]) di er from STRIPS in the knowledge representations and reasoning methods they use. But they too share the core assumption that qualitative methods are largely su cient for navigational planning in most, if not all, spatial worlds. Traditional robotics techniques for spatial navigation too rely on deliberative planning. But in contrast to AI methods, robotics techniques for navigational planning are based on numerical representations and reasoning [Latombe 1991]. This is because, the argument runs, the movement of a robot needs to be computed with an accuracy and precision that is beyond qualitative representations and reasoning: accurate navigation from an initial location to a goal location requires numerical measures such as the precise distance the robot needs to traverse and the precise direction in which it needs to travel. Recent work on robotics (e.g., [Brooks 1986]) shifts the emphasis and focus from navigational planning to situated action. This line of research posits mechanisms for reactive control that directly map perceptions of the world to actions on it. The perceptual inputs, motor outputs, and internal mappings, all are numerical. But there are no internal representations of the world, qualitative or numerical, and hence no deliberative planning either: the world is its own best representation. The development of reactive-control techniques in turn has given rise to hybrid robot architectures that combine navigational planning and reactive control. The AuRA architecture [Arkin 1989a], for example, uses navigational planning to form a high-level plan for achieving a given goal and reactive control to avoid obstacles not anticipated by the planner. Following traditional robotics techniques, navigational planning in hybrid robot architectures typically uses numerical spatial representations and reasoning. This sets up the research question for our work: Given a hybrid robot architecture capable of both navigational planning and reactive control, might qualitative representations and reasoning su ce for navigational planning? Or, are numerical representations and methods really needed for navigational planning even in a hybrid architecture? This issue is important because it pertains to the nature of spatial representations and reasoning for navigational planning. To explore this issue we formulated a testable (i.e., falsi able) hypothesis: For hybrid robots capable of both navigational planning and reactive control, qualitative representations and reasoning are su cient for navigational planning. The characterization of what we mean by \su cient" will emerge in the following sections and is made explicit in the last section. But note that optimality of planning or of navigation is not an issue here; neither qualitative-planningmethods nor reactive-control techniques provide any a priori guarantee of planning or navigational optimality, especially not in spatial worlds containing obstacles in addition to the goal. The engineering of an optimal planner or navigator is not the goal of our work. Instead, our goal is only to examine and evaluate a speci c hypothesis of considerable intrinsic merit. If this hypothesis is false, then qualitative methods are insu cient for all classes of spatial worlds (barring toy or imaginary worlds, of course). But if the hypothesis is true, then qualitative methods are su cient for navigating in at least one class of spatial worlds. Evaluation Method and Experimental