Online Action Recognition

Recognition in planning seeks to find agent intentions, goals or activities given a set of observations and knowledge library (e.g. goal states, plans domain theories). In this work we introduce the problem Online Action Recognition. It consists recognizing, an open world, action that best explains partially observable state transition from first-order STRIPS actions, which is initially empty. We frame as optimization problem, propose two algorithms address it: Unification (AU) through (OARU). The former builds on logic unification generalizes input actions using weighted partial MaxSAT. latter looks for within observed transition. If there such action, it making use AU, building way AU hierarchy. Otherwise, OARU inserts Trivial Grounded (TGA) just report results benchmarks International Planning Competition PDDLGym, where recognizes accurately with respect expert knowledge, shows real-time performance.