Role Transfer for Robot Tasking
MASSACHUSETTS INST OF TECH CAMBRIDGE DEPT OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE
Pagination or Media Count:
Robotics has proven most successful in narrowly defined domains that offer sufficient constraints to make automated perception and action tractable. The goal of this thesis is to take a step towards generality by developing methods for applying a robot to many different narrow domains. This is complementary to the more common research goal of enhancing machine perception and action to deal with wider domains. This approach to extending the range of application of a technology through parameterization rather than generalization is key to fields such as automatic speech recognition. It has the theoretical advantage of providing a framework for factoring context into perception, and the practical advantage of creating systems that do useful work with limited technology. I propose a scheme for communicating constraints to a mechanically general-purpose robot, so that it can perform novel tasks without needing to first solve open problems in perception and action. In particular, this thesis develops mechanisms for communicating the structure of simple tasks to a robot, translating this structure into a set of supervised learning problems for parts of the task which are difficult to communicate directly, and solving those problems with the guidance of a protocol for inducing feature selection.