Research Planning Grants (RPG): Intelligent Robot Control

小行星9407363 神经网络利用大量传感器信息、并行处理、学习和泛化的能力使其成为智能系统控制的有吸引力的候选者。 本研究的目的是设计并实现一个工业机器人的智能神经控制器。 五个基本方案已被用于机器人控制的神经网络监督控制，直接逆，神经自适应，反向传播的效用，自适应的批评。 虽然前三种方法对于跟踪控制是好的，但真正的智能控制系统应该将系统作为一个整体来考虑，并且应该包括规划和性能优化。 为此，反向传播的效用和自适应评论家计划具有最大的潜力，因为他们最大限度地提高性能的基础上的模型或预测未来的利用率。 本研究将探讨三个机器人系统，2-D平面机器人，3-D斯坦福大学的手臂，和6-D Scorbot ER-III机器人的优势和劣势的反向传播的实用程序和自适应评论计划。 这些包括仿真，评估，并最终建立机器人控制的设计准则。 通过这三个系统的集成，从平面运动到复杂的六度运动，该项目将试图为机器人系统的神经控制提供完整而系统的研究。

9407363 Tang A Neural network's ability to utilize large amounts of sensory information, parallel processing, learning and generalization has made it an attractive candidate for intelligent system control. The objective of this research is to design and implement an intelligent neuro-controller for industrial robots. Five basic schemes have been used for robot control based on neural networks supervised control, direct inverse, neural adaptive, backpropagation of utility, and adaptive critics. While the first three methods are good for tracking control, a truly intelligent control system should consider the system as a whole and should include planning and performance optimization. Towards this end, the backpropagation of utility and adaptive critic schemes have the greatest potential because they maximize performance based on models or predictions of future utilization. This research will investigate the strengths and weakness of the backpropagation of utility and adaptive critic schemes for three robotics systems, 2-D planar robot, 3-D Stanford arm, and the 6-D Scorbot ER-III robot. These include simulation, evaluation, and ultimately the establishment of design guidelines for robot control. With the integration of these three systems, from planar motion to complex six-degree motion, this project will attempt to provide a complete and systematic study of neuro-control for robotics systems.