Online Recognition of Navigation Goals Through Goal Mirroring

Goal recognition is the problem of inferring the (unobserved) goal of an agent, based on a sequence of its observed actions [4, 2, 1, 5]. It is a fundamental research problem in artificial intelligence, closely related to plan, activity, and intent recognition [10]. In offline recognition the entire sequence of observations is provided to the agent ahead of time. In contrast, in online recognition the sequence of observations is revealed incrementally instead of being known in advance, thus exacerbating an already hard problem. The prevalent approach to goal recognition, both offline and online, relies on a dedicated plan library, a set of plans which represents all known ways to achieve known goals [10]. These recognition methods vary in the expressiveness of the representation and efficiency of the inference algorithms used. While powerful when the plans are known, these methods fail when the observations come from an unknown plan to achieve a known goal. An additional difficulty is raised when adding goals to the set of recognizable goals, as plans for them need to be inserted in the library, in order to be recognized. One of the notable exceptions is plan recognition by planning (PRP), which focuses on library-free recognition, where a planner is used as a black box, to dynamically generate plans that are matched against the observations, eliminating the need for a plan library [6, 8]. This approach targets discrete domains only, and is inefficient for online recognition where it would produce 2|O||G| calls to the planner, where |O| is the number of observations, and |G| the number of goals. We advocate goal mirroring. Like [6], goal mirroring uses a planner, to generate recognition hypotheses. However, it is designed for efficient, online recognition in continuous environments by using a motion planner, with a baseline number of calls to the planner of (|O|+ 1)|G|. We additionally identify two key decision points where, by inserting heuristics for navigational goal recognition, we can further influence the number of calls to the planner and the overall run-time.