Planning with nondeterministic actions and sensing

Many planning problems involve nondeterministic actions-actions whose effects are not completely determined by the state of the world before the action is executed. In this pa-per we consider the computational complexity of planning in domains where such actions are available. We give a for-mal model of nondeterministic actions and sensing, together with an action language for specifying planning domains. Then, we examine the cases of complete observability, par-tial observability and no observability, assuming that sensing is done automatically or needs to be done explicitly. We re-strict our attention to plans of tractable plan-size or depth. We show that planning with nondeterministic actions for polynomially represented plans has computational complex-ity equivalent to that of planning with deterministic actions under incomplete knowledge about the initial state, if we the domains include no observability or full observability, and consider an assumption on executability of actions. If the latter takes polynomial time, then our complexity class for all these problems is -complete. If the problem of check-ing executability of actions is NP-complete (the general case), or we allow partial observability or sensing actions, then our complexity class is-complete. For plans of polynomial depth, we find that planning in nondeterministic systems with no observations is in -complete, contrary to previous con-jectures of PSPACE-completeness (Haslum & Jonsson 1999). We also find that planning in nondeterministic systems with full observability or partial observability (with and without sensing actions) is PSPACE-complete for polynomial-depth plans. These results point out cases where it may be useful to use encodings in boolean formulae to perform planning, and they carefully draw the distinctions between the different scenarios involved.