Scalable Inference of Affordance, Activity, and Intent from Spatio-Temporal Input
Abstract
The proposed research seeks to develop architectures for robots that, given streams of noisy perceptual inputs and incomplete domain"" knowledge, enable plausible and scalable inference about other agents~ activities, intentions, and affordances. Towards this object""ive, we will develop novel representational ideas, and two systems that operate over these ideas. Specifically, we will make the fol""lowing three primary contributions:1. A hierarchical representation for activities and concepts, with the ability to represent def""ault knowledge, support distributed encoding, represent affordances as relations defined jointly over objects and agents, and to ass""ociate probabilities with concepts and activities.2. A system that will combine the capabilities of action languages, declarative"" programming, and probabilisticreasoning, to support inference with an incomplete system description, violation of defaults, noisy"" observations of system behavior, and minimal use of heuristics.3. A system that will draw inferences in an incremental, data-driv""en form of abductive inference guided by heuristic rules, interpolating observations and background information with domain axioms t"hat are ground dynamically. These contributions will be thoroughly evaluated in simulation and on physical robots in scenarios that mimic the task complexity and spatial context of scenarios of naval relevance. We will analyze the effects of factors such as comp"lexity (of the knowledge base and the explanation generated) and sensor reliability, using accuracy and computational efficiency as" the performance measures.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2017
- Source ID
- N000141712434
Entities
People
- Mohan Sridharan
Organizations
- Office of Naval Research
- United States Navy
- University of Auckland