Embodied Computational Metacognition
Abstract
While humans can rapidly expand their vocabulary of concepts with few or no examples, and generalize from previous to new experiences, modern machine learning techniques do not easily or organically expand to accommodate new concepts. This project proposes to combine the relative strengths of three common AI approaches: 1) Machine learning, 2) Embodied simulation, and 3) Logical reasoning, to develop agents that rapidly detect changes in their environments as they operate within them, and acquire physically-grounded conceptual representations to interpret novel classes of objects and events they encounter. This capacity will facilitate computational "fast mapping" as is observed in certain human cognitive processes. A three-year effort is proposed to develop methods by which an agent exploring an environment can use the geometric and spatial properties of objects it interacts with along with the semantics of the events it executes, to detect when the environment itself has changed and the agent needs to expand its conceptual model, e.g., to accommodate the introduction of a novel object, event, or a change in the situational context. Agents will explore simulated worlds and interact with objects in the course of solving tasks that require making inferences about circumstances that have never been previously encountered, and making analogies to known phenomena in order to define novel information. Because newly acquired concepts must be grounded to previously known concepts and characteristics of the environment and interaction, a successful agent will be able to explain its reasoning for changing its model, and its interpretation of phenomena that are novel to it. We propose 3 tasks to address the objectives of situationally-aware metacognitive agents 1) Detecting Novel Phenomena; 2) Analogical Task Reasoning; and 3) Explaining Context-Dependent Decisions. Thus if successful, this work will contribute to agents that can both effectively team with humans and operate semi-independently with situational awareness in the field. Such agents can do more with less guidance from humans, meaning they can be force multipliers in situations dangerous or inaccessible to humans, and can continue operations in communication-denied environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 19, 2023
- Source ID
- W911NF2310031
Entities
People
- Nikhil Krishnaswamy
Organizations
- Army Contracting Command
- Colorado State University
- United States Army