AbstractsComputer Science

Intelligent collaborative agents are becoming common in the human society. From virtual assistants such as Siri and Google Now to assistive robots, they contribute to human activities in a variety of ways. As they become more pervasive, the challenge of customizing them to a variety of environments and tasks becomes critical. It is infeasible for engineers to program them for each individual use. Our research aims at building interactive robots and agents that adapt to new environments autonomously by interacting with human users using natural modalities. This dissertation studies the problem of learning novel tasks from human-agent dialog. We propose a novel approach for interactive task learning, situated interactive instruction (SII), and investigate approaches to three computational challenges that arise in designing SII agents: situated comprehension, mixed-initiative interaction, and interactive task learning. We propose a novel mixed-modality grounded representation for task verbs which encompasses their lexical, semantic, and task-oriented aspects. This representation is useful in situated comprehension and can be learned through human-agent interactions. We introduce the Indexical Model of comprehension that can exploit extra-linguistic contexts for resolving semantic ambiguities in situated comprehension of task commands. The Indexical model is integrated with a mixed-initiative interaction model that facilitates a flexible task-oriented human-agent dialog. This dialog serves as the basis of interactive task learning. We propose an interactive variation of explanation-based learning that can acquire the proposed representation. We demonstrate that our learning paradigm is efficient, can transfer knowledge between structurally similar tasks, integrates agent-driven exploration with instructional learning, and can acquire several tasks. The methods proposed in this thesis are integrated in Rosie - a generally instructable agent developed in the Soar cognitive architecture and embodied on a table-top robot.