The use of coreference resolution for understanding manipulation commands for the PR2 Robot

Abstract

Natural language interaction can enable us to interface with robots such as the Personal Robot 2 (PR2), without the need for a special training or equipment. Programming such a robot to follow commands is challenging because natural language has a complex structure and semantics, a model for which needs to be based on linguistic knowledge or learned from examples. In this thesis we first enable the PR2 robot to follow manipulation commands expressed in natural language by applying the Generalized Grounding Graph (G3 ). We model the PR2's actions and their trajectories in the physical environment, define the state-action space and learn a grounding model from an annotated corpus of robot actions aligned with commands. We achieved lower overall performance than previous implementations of G3 had reported. After that, we present an approach for using the linguistic technique of coreference resolution to improve the robot's ability to understand commands consisting of multiple clauses. We constrain the groundings for coreferent phrases to be identical by merging their nodes in the grounding graph. We show that using coreference information increases the robot ability to infer the right action sequence. This brings the robotic capabilities of modeling and understanding natural language closer to our theoretical understanding of discourse.