TYRO: A robot vision system that can learn from observations of its environment
Abstract
The ability to learn from experience is one of the hallmarks of intelligent behavior. If a robotic system is to be considered as truly intelligent, then it must be able to learn from its experiences. What is more, since an intelligent robotic system must interact with its environment, rather than merely act upon it in a blind fashion, then it should be able to learn from its observations. Vision seems to be a most powerful sense, and so we have concentrated on the problem of learning from visual observations. The first part of this thesis details methods for gathering and interpreting information from a range sensor in order to produce a scene description. These methods are accurate even when the objects are of irregular shapes and the sensory information is noisy, and have been implemented in our vision system P scILE-I scDENTIFY. In the second part, we turn our attention to the problem of learning. Given a set of object descriptions, a set of expressions that specify measurements of specific instances to which a human supervisor has assigned class identities, an intelligent system should be able to discover expressions that can serve as class descriptions. We would like for these expressions to describe the classes accurately, so that the description of class X is likely to hold for future instances of that class, instances to which the supervisor would assign the class identity X, although the system has not yet observed them. At the same time, it should be unlikely that the expression would also hold for future instances of other classes. In addition to accuracy, we would also like for the expressions to be meaningful even to a naive observer. Our system T scYRO builds on symbolic learning techniques to meet both these goals in practical applications that include biomedical image analysis and both 2-D and 3-D machine vision.
Degree
Ph.D.
Advisors
Kak, Purdue University.
Subject Area
Electrical engineering|Artificial intelligence|Industrial engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.