Goal recognition is generally considered to follow plan recognition. The plan recognition problem is typically defined to be that of identifying which plan in a given library of plans is being executed, given a sequence of observed actions. Once a plan has been identified, the goal of the plan can be assumed to follow. In this work, we address the problem of goal recognition directly, without assuming a plan library. Instead, we start with a domain description, just as is used for plan construction, and a sequence of action observations. The task, then, is to identify which possible goal state is the ultimate destination of the trajectory being observed.

We present a formalisation of the problem and motivate its interest, before describing some simplifying assumptions we have made to arrive at a first implementation of a goal recognition system, AUTOGRAPH. We discuss the techniques employed in AUTOGRAPH to arrive at a tractable approximation of the goal recognition problem and show results for the system we have implemented.