REGENTS OF THE UNIVERSITY OF MICHIGAN Ann Arbor United States
Our research on interactive task learning has emphasized the continued development of a natural language understanding system that interfaces with a human instructor, and the underlying task learning system. The language system uniquely combined relevant research a construction grammar approach to representing linguistic knowledge an incremental, single-path processing algorithm with local repair a cognitive architecture as the computational platform and embodiment in the robotic agent for grounding language to the agents perception, action capabilities, and knowledge of the world. We extended this to process real world task instructions and the ability to handle multiple forms of ambiguity. The task learning research led to the development of an agent for learning all aspects of tasks, with emphasis on handling ambiguous scenarios. The agent can interactively learning over 55 games and puzzles, and transfers knowledge learned in one game to a similar game.