Model Based Neural Networks for Kinematic Control of Robots  

ABSTRACT

Using robot arms, or a collection of them, to perform various tasks is becoming increasingly popular in both industry and our daily life. Recent advances in machine learning provide us with an opportunity to employ innovative learning to reach autonomous control. Most existing neural network based robot control methods are designed based on statistical criteria with reasonable average performance, but lacks a mechanism to ensure worst-case performance. This talk will introduce our research on constructing neural networks based on knowledge of the robot model and apply it to solve robot arm manipulation problems. This approach inherits the advantage of neural networks in adaption and satisfies guaranteed stability. It provides a tractable solution using neural networks to address safety critical tasks with certified performance.