Inferring beliefs for search and rescue from natural language

Abstract

A learned natural language robotic interface can allow a human operator to intuitively communicate instructions to a robot. A number of models, including probabilistic grounding graphs, have been used to ground natural language input to the real-world tasks a robot must perform. In this thesis, I provide two extensions to existing work in grounding natural language instructions. First, I apply an existing probabilistic grounding graph model in the context of outdoor search and rescue, introducing a new set of groundings to allow a continuous cost map to be inferred from the natural language. Second, I incorporate pool-based active learning into the training of the probabilistic grounding graph model, which shows promise for reducing the number of labeled examples needed to train the model.