tim@LINC.CIS.UPENN.EDU (Tim Finin) (03/05/87)
Dissertation Defense
Computer and Information Science
University of Pennsylvania
Real Time Expert System to Control a Robot Ping Pong Player
R.L. Andersson
A real time "expert" control system has been designed
and forms the nucleus of a functioning robot ping pong
player.
Robot ping pong is underconstrained in the task
specification (hit the ball back), and heavily constrained
by the manipulator capabilities. The expert system must
integrate the sensor data, robot capabilities, and task
constraints to generate an acceptable plan of action. The
robot ping pong task demands that the planner anticipate
environmental changes occurring during planning and robot
motion. The inability to generate accurate, timely plans
even in the face of a capricious environment and limited
actuator performance would result in a nonfunctional system.
The program must continuously update the task plan as
new sensor data arrives, selecting appropriate modifications
to the existing plan, rather than treating each datum
independently. The difficult task and the stream of sensor
data result in an interesting system architecture. The
expert system operates in the symbolic and numeric domains,
with a blackboard to enable global optimization by local
agents. The architecture interrelates initial planning,
temporal updating, and exception handling for robustness.
A sensor and processing system produces three
dimensional position, velocity, and spin vectors plus a time
coordinate at 60 Hz. Novel processing algorithms and
careful attention to camera modeling were necessary to
obtain adequate accuracy.
A robot controller provides accurate, predictable
performance close to the envelope of robot capabilities
using modeling and feed-forward techniques. The controller
allows motions to be planned in the temporal domain
including specified terminal velocities, and supports smooth
changes to motions in progress.
Performance of the sensor subsystem, actuator and robot
controller, and expert system will be demonstrated. The
system successfully plays against both human and machine
opponents.
COMMITTEE: DR. LOU PAUL (ADVISOR)
DR. TIM FININ
DR. RUZENA BAJCSY
DR. ROD BROOKS (MIT)
DATE: MARCH 27, 1987
TIME: 10-12 NOON
LOCATION: 129 MOORE (FACULTY LOUNGE)