Neural Network Controller for Adaptive Hand-Eye Coordination in Robots

摘要

The overall objective of Phase I work was to implement a prototype of a neural network controller that adaptively coordinates a visually-guided multijoint arm to reach elongated objects in space. The adaptive controller can learn to accurately grasp an elongated object without any information about the geometry of the robot-camera system. The neural controller learns about space by associating how the arm moves with what the cameras see. Learning is achieved without any external source of error correction; this allows the controller to be completely independent of supervision. The implementation is fault tolerant to hardware errors. The neural controller adapts to unforeseen changes in the geometry of the robot system and the shape and size of the object. The system will self-organize and maintain motor calibrations starting with only loosely defined relationships. The system adaptability replaces the need for operator calibration.