Asking Human Help in Contingent Planning

Contingent planning models a robot in a partially observable environment and (non)deterministic actions. In a contingent planning problem, a solution can be found by doing a search in a space of belief states, where a belief state is represented by a set of possible states. However, in the presence of deadend belief states, a situation where a robot may fail to complete its task, the only way to accomplish it is to ask for human help. In this work, rather than limiting a contingent planning task to those that only include actions and observations that an agent can perform autonomously, agents can instead reason about asking humans in the environment for help in order to complete tasks that would be unsolvable otherwise. We are interested in develop agents with symbiotic autonomy: an agent that proactively and autonomously ask for human help. We make the assumption that humans may only be interrupted when it is extremely necessary, i.e. when the planner can not solve a task or solving it implies in a too high cost. To solve this problem we propose an extension of an existing translation technique [2] that translates a contingent planning problem into a non-deterministic fully observable planning problem. The proposed translation can deal with different types of dead-end belief states, including a pure dead-end (e.g. a broken robot) and a dead-end that is due to uncertainty about the environment.