Doctoral Thesis: Mutual Information-based Gradient-ascent Control for Distributed Robotics

SHARE:

Event Speaker: 

Brian J. Julian

Event Location: 

32-G449 (Kiva)

Event Date/Time: 

Tuesday, July 23, 2013 - 10:00am

Abstract:

In this defense I present the derivation, analysis, and implementation of a novel class of decentralized mutual information-based gradient-ascent controllers.  Built on an information-theoretic foundation, these controllers use the analytical gradient of mutual information to continuously move robots equipped with sensors to better observe their environment.  I first show that maximizing the mutual information between the environment state and the robots’ next joint observation is relevant to many inference tasks such as occupancy grid mapping.  I then prove the correctness of these controllers; given simple robot dynamics and few probabilistic assumptions, these controllers are convergent between sensor observations and, in their most general form, locally optimal.   Lastly, by employing distributed approximation algorithms and non-parametric sample-based methods, I show that these controllers scale to large multi-robot teams.  More specifically, the computational complexity of these controllers is independent of the number of robots. Throughout the defense I support my work with numerical simulations and hardware experiments concerning traditional robot applications such as mapping, exploration, and surveillance.

Bio:

Brian J. Julian received the B.S. degree in mechanical and aerospace engineering from Cornell University, and the S.M. degree in electrical engineering and computer science from the Massachusetts Institute of Technology.  He is currently a Ph.D. candidate under Prof. Daniela Rus in MIT CSAIL's Distributed Robotics Laboratory. Since 2005, he has been a staff member in the Engineering Division at MIT Lincoln Laboratory, where his is currently Associate Staff in the Rapid Prototyping Group and a Lincoln Doctoral Scholar.  Brian’s research in distributed multi-robot coordination applies tools from control, network, and optimization theory to develop scalable information-theoretic algorithms with provable performance characteristics.
 
Thesis Committee:  Prof. Daniela Rus (advisor), Prof. Leslie Kaelbling, Prof. John Leonard, and Prof. Mac Schwager