The biggest challenges facing robotics today are issues dealing with novelty, processor faults and self-organization. To meet these needs, the proposed study will implement a prototype of a neural network model used to adaptively control the coordination of a multijoint robot arm and two stereo cameras. The implementation will be based on the computer simulations of an existing neural network model of adaptive hand-eye coordination. The implementation will evolve smoothly in two generations: the first, with commercially available equipment and the second, with new dedicated hardware. In phase I of the SBIR work, the implementation will be designed for a combination of image processing hardware and a robot arm. When fully implemented on dedicated hardware in SBIR phase II, this prototype system will coordinate a multi-joint robot arm to adaptively reach targets in three dimensions in real time. The neural system self-organizes and maintains visual- motor calibrations starting with only loosely defined relationships. The system tolerates internal noise, partial system damage and changes in the mechanical and optical parameters of the robot as they occur during wear. This adaptability will replace operator calibration in many different robots.
