Toward Undifferentiated Cognitive Agents: Translating Instructions to Knowledge
Abstract
In this project, we are developing a cognitive system capable of independently acquiring most required task knowledge and skill to perform a set of tasks. Our proposed approach begins with the development of a foundational and generalizable cognitive system that can be transformed into a specialized cognitive agent through written instruction, interactions with its trainers, task experience, and developer intervention (when needed). In this vein, we have developed an undifferentiated agent (uAgent) that is a set of general-purpose computational cognitive capacities enabling it to read task instructions, iteratively interact with trainers to fill gaps in task knowledge, generate the requisite task knowledge from the instructions, and complete multitasking scenarios of varying complexity. The project represents a collaboration among several institutions, including Drexel University, Kansas State University, and the Air Force Research Laboratory. The Drexel portion of the project focuses specifically on the problem of translating instructions to knowledge. Within this effort, we have developed several key contributions: (1) a unified framework within a "cognitive code" cognitive architecture that encodes realistic instructions and translates them to executable knowledge; (2) within the framework, an approach to address key challenges related to interpreting English-like instructions, such as anaphora resolution, grounding, and interactive instruction.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 22, 2023
- Accession Number
- AD1231002
Entities
People
- Dario D Salvucci
Organizations
- Drexel University