Contingent Planning in Domains with Unsafe Facts

Contingent planning problems model agents with partial information about their state, that can use sensing actions to observe and reason about unknown values. Such problems can be efficiently solved using an online approach, where the agent replans each time new information is obtained through sensing.In this paper, we study Unsafe PPOS (U-PPOS) problems, where the input specification may contain unsafe facts—facts whose assumed value in the model differs from reality. Specifically, we consider false positives (FPs), which are assumed true but are actually false, and false negatives (FNs), which are assumed false but are actually true. FPs may lead planners to action failures, while FNs may cause the agent to believe that deadends exist. We propose methods to detect and handle both types of inaccuracies, establish their theoretical properties, and evaluate their performance across a set of benchmark domains.