Generating Behavior: Would Custom Hardware Help?

Prof. Chris Atkeson
CMU School of Computer Science

Abstract: This talk will discuss some current approaches to generating
behavior with the goal of exploring whether there are computational
hardware features that might make behavior generation faster and more
effective.  Behavior generation is an important issue in applications
including robotics, video games, and perception. One current approach
to behavior generation is selecting behaviors from a library of stored
experience.  This memory/case/exemplar/instance-based approach relies
on rapid nearest neighbor search of large databases. Another approach
to behavior generation is optimization, in which large constrained
nonlinear function minimization problems have to be solved in real
time, or solutions to anticipated problems created and stored in
advance.

Bio: Chris Atkeson is a Professor in the Robotics Institute and
Human-Computer Interaction Institute at CMU.  He focuses on humanoid
robotics and robot learning.  He has enabled humanoid robots to learn
to perform various dynamic tasks such as juggling.  Chris has done
work in local learning, learning from demonstration, learning from
practice, memory-based learning, and reinforcement learning.  More
information and publications are available from his web page
www.cs.cmu.edu/~cga.