Context-Rich Predictors for Self-Reflective Autonomy: Variational Foundations
Abstract
Among the numerous challenges in autonomy, many are centered around the ability to estimate the current and future state of the world, especially in new and unstructured environments. It is clear that simply relying on the current sensor data alone has severe limitations.Sensors may be too noisy, capture only parts of the environment, and say little about the future. Success of humans and animals to operate in unfamiliar and unstructured environments is due, in part, to two capabilities. First, we are able to predict the current state and even the distant future using contextual information, usually derived from experiences, as a supplement to limited sensor data. Second, we have the ability to assess the effectiveness of a prediction protocol and make adjustments when necessary. The importance of contextual information is of course widely recognized and major efforts in this domain are underwaysuch as those at the Contextual Robotics Institute, UC San Diego. In this proposal, we offer to develop new mathematical tools that will improve the predictive capability of autonomous systems and their ability to self-reect on such capabilities. Specifically, we focus on twomajor tasks:Major Task 1: Develop and analyze a broad class of variational-based nonparametric estimators and ?lters (context-rich predictors) that, in addition to sensor data, account for experiences, contextual information, and uncertainty in such information.Major Task 2: Develop variational theory and protocols for assessing, selecting, constructing, and instantiating context-rich predictors based on sensor data, experiences, and contextual information, and thereby support the development of self-reective autonomy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2017
- Source ID
- N000141712372
Entities
People
- Roger J-B Wets
Organizations
- Office of Naval Research
- United States Navy
- University of California, Davis