Integrating human-provided information into belief state representation using dynamic factorization

Abstract

In partially observed environments, it can be useful for a human to provide the robot with declarative information that augments its direct sensory observations. For instance, given a robot on a search-and-rescue mission, a human operator might suggest locations of interest. We provide a representation for the robot's internal knowledge that supports efficient combination of raw sensory information with high-level declarative information presented in a formal language. Computational efficiency is achieved by dynamically selecting an appropriate factoring of the belief state, combining aspects of the belief when they are correlated through information and separating them when they are not. This strategy works in open domains, in which the set of possible objects is not known in advance, and provides significant improvements in inference time, leading to more efficient planning for complex partially observable tasks. We validate our approach experimentally in two open-domain planning problems: a 2D discrete gridworld task and a 3D continuous cooking task.