The field of AI planning has seen rapid progress over the last decade and planners are now able to find plan with hundreds of actions in a matter of seconds. Despite those important progresses, robotic systems still tend to have a reactive architecture with very little deliberation on the course of the plan they might follow. In this thesis, we argue that a successful integration with a robotic system requires the planner to have capacities for both temporal and hierarchical reasoning. The former is indeed a universal resource central in many robot activities while the latter is a critical component for the integration of reasoning capabilities at different abstraction levels, typically starting with a high level view of an activity that is iteratively refined down to motion primitives. As a first step to carry out this vision, we present a model for temporal planning unifying the generative and hierarchical approaches. At the center of the model are temporal action templates, similar to those of PDDL complemented with a specification of the initial state as well as the expected evolution of the environment over time. In addition, our model allows for the specification of hierarchical knowledge possibly with a partial coverage. Consequently, our model generalizes the existing generative and HTN approaches together with an explicit time representation. In the second chapter, we introduce a planning procedure suitable for our planning model. In order to support hierarchical features, we extend the existing Partial-Order Causal Link approach used in many constraintbased planners, with the notions of task and decomposition. We implement it in FAPE (Flexible Acting and Planning Environment) together with automated problem analysis techniques used for search guidance. We show FAPE to have performance similar to state of the art temporal planners when used in a generative setting. The addition of hierarchical information leads to further performance gain and allows us to outperform traditional planners. In the third chapter, we study the usual methods used to reason on temporal uncertainty while planning. We relax the usual assumption of total observability and instead provide techniques to reason on the observations needed to maintain a plan dispatchable. We show how such needed observations can be detected at planning time and incrementally dealt with by considering the appropriate sensing actions. In a final chapter, we discuss the place of the proposed planning system as a central component for the control of a robotic actor. We demonstrate how the explicit time representation facilitates plan monitoring and action dispatching when dealing with contingent events that require observation. We take advantage of the constraint-based and hierarchical representation to facilitate both plan-repair procedures as well opportunistic plan refinement at acting time.
展开▼
