Karen Haigh, Research Interests

Degrees: B.Sc. (U. of Ottawa, Honours computer science, summa cum laude)

Year entered CMU: 1992

Advisor: Manuela M. Veloso

Research Interests: Machine Learning, Robotics, Case-based Reasoning,
                    Planning, PRODIGY

===============================================================================
                Using Planning and Execution Experience for
                         High-Level Robot Learning

                                Thesis Proposal
                                 December 1995

                               Karen Zita Haigh
                          School of Computer Science
                          Carnegie Mellon University
                             Pittsburgh, PA 15213

     Robots  have rather  limited   capabilities for  high-level  reasoning
     where  planning  for complex tasks  is  required.  As a result, recent
     years  have seen   a growing interest    in  systems which have   both
     planning  capabilities for  providing goal-directedness  and  reactive
     runtime capabilities for providing flexibility  and sensitivity to the
     environment. However, it is not  sufficient for an autonomous agent to
     simply have goal-directedness and reactive capabilities;  it must have
     the  ability   to  learn.  Without   learning,  the  behaviour  of  an
     autonomous agent is completely dependent on  the predictive ability of
     the programmer.  To be truly autonomous, the agent needs to be able to
     use  accumulated  experience  and feedback about  its  performance  to
     improve its behaviour.

     I  propose in  my thesis work   to investigate  learning mechanisms to
     improve  high-level reasoning in autonomous robots.  I  will use as my
     platform  Rogue, the integrated  architecture I  have built  on top of
     PRODIGY, a  planning  and learning system,  to control  the high-level
     task planning for Xavier, an autonomous indoor mobile robot.  Rogue is
     intended to  be a roving office gofer unit,  and  will deal with tasks
     such  as delivering mail,  picking  up printouts and returning library
     books.

     My first objective in this work is to identify when and where failures
     are likely to occur and use that knowledge to avoid them and to create
     contingency plans.  Second,  I propose  to build a  mechanism  to form
     performance expectations  which can  be used  to place constraints  on
     tasks and  make predictions about the completion  of tasks. Finally, I
     intend to implement a quality feedback mechanism  so Rogue can improve
     its definitions of job priority and task compatibility, which are used
     to determine the order in which tasks are  completed and when they can
     be successfully merged. I expect  the major contributions of this work
     to be: an implementation of a planning,  executing and learning system
     on a real  robot; an  improvement in  the reliability, usability   and
     quality of robot plans;   and the emergence of   synergistic learning,
     where the various learning mechanisms aid each other.

     In summary,  my system will learn from real-world execution experience
     to improve its high-level reasoning capabilities.