Fast and Informed Action Selection for Planning with Sensing

Abstract

Consider a robot whose task is to pick up some colored balls from a grid, taking the red balls to a red spot, the blue balls to a blue spot and so on, one by one, without knowing either the location or color of the balls but having a sensor that can find out both when a ball is near. This problem is simple and can be solved by a domain-independent contingent planner in principle, but in practice this is not possible: the size of any valid plan constructed by a contingent planner is exponential in the number of observations which in these problems is very large. This doesn’t mean that planning techniques are of no use for these problems but that building or verifying complete contingent plans is not feasible in general. In this work, we develop a domain-independent action selection mechanism that does not build full contingent plans but just chooses the action to do next in a closed-loop fashion. For this to work, however, the mechanism must be both fast and informed. We take advantage of recent ideas that allow delete and precondition-free contingent problems to be converted into conformant problems, and conformant problems into classical ones, for mapping the action selection problem in contingent planning into an action selection problem in classical planning that takes sensing actions into account. The formulation is tested over standard contingent planning benchmarks and problems that require plans of exponential size.