Backward Plan Construction under Partial Observability

We present algorithms for partially observable planning that iteratively compute belief states with an increasing distance to the goal states. The algorithms handle nondeterministic operators, but restrict to problem instances with a finite upper bound on execution length, that is plans without loops. We discuss an implementation of the algorithms which uses binary decision diagrams for representing belief states. It turns out that generation of new belief states from existing ones, corresponding to the use of conditional branches in the plans, can very naturally be represented as standard operations on binary decision diagrams. We also give a preliminary experimental evaluation of the algorithms.