Robotic hand-eye motor learning

Citation

Ruoff, Carl F. (1993) Robotic hand-eye motor learning. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/47t1-2f31. https://resolver.caltech.edu/CaltechETD:etd-12052007-130729

Abstract

This thesis investigates the use of neural networks and nonlinear estimation in robotic motor learning. It presents a detailed experimental investigation of the performance and parametric sensitivity of resource-allocating neural networks along with a new learning algorithm that offers rapid adaptation and excellent accuracy. It also includes an appendix that relates feed-forward neural networks to familiar mathematical ideas. In addition, it presents two learning hand-eye calibration systems, one based on neural networks and the other on nonlinear estimation. The network-based system learns to correct robot positioning errors arising from the use of nominal system kinematics, while the estimation-based system identifies the robot's kinematic parameters. Both systems employ the same two-link robot with stereo vision, and include noise and various other error sources. The network-based system is robust to all error sources considered, though noise naturally limits performance. The estimation-based system has significantly better performance when the robot and vision systems are well modeled, but is extremely sensitive to unmodeled error sources and noise. Finally, it presents a robot control system based on neural networks that learns to catch balls perfectly without requiring explicit programming or conventional controllers. It uses only feed-forward pursuit motions learned through practice, and is initially incapable of even moving its arm in response to external stimuli. It learns to identify and control its pursuit movements, to identify and predict ball behavior, and, with the aid of advice from a critic, to modify its movement commands to improve catching success. The system, which incorporates information from visual, arm state, and drive force sensors, characterizes control situations using input/response pairs. This allows it to learn and respond to plant variations without requiring parametric models or parameter identification. It achieves robust execution by comparing predicted and observed behavior, using inconsistencies to trigger learning and behavioral change. The architectural approach, which involves both declarative and analog knowledge as well as short- and long-term memory, can be extended to learning other sensor-motor skills like mechanical assembly and synchronizing motor actions with external processes.

Thesis Files