Data-driven, adaptive control of servo drives for industrial robots

Abstract

Data-driven estimation algorithms (i.e., artificial neural networks or support vector machines) are a robust alternative to conventional, physical models for adaptive controllers. These algorithms can identify non-linear and time-variant dynamic systems and predict the systems’ future outputs using empirical data. Thus, deep knowledge of the system’s physics is not required anymore, and models can be implemented and derived easily and quickly. Especially for industrial robots, data-driven estimation algorithms offer new possibilities. For example, depending on the pose and load, the non-linear dynamics of the robot do not need to be calculated by means of complex kinematic equations but can be estimated based on sensor data. This saves computational time and provides a constant model accuracy. Furthermore, the controller can adapt to changed boundary conditions, such as varying mechanical loads on the end effector. This paper discusses the design of a data-driven, model predictive controller for a six- and a single-axis joint test rig. Here, the design and validation of the controller are based on a co-simulation consisting of a multibody and system simulation. All design parameters of the control (order of magnitude of the model, training data, optimization, etc.) are discussed. Finally, the control will be validated in an experiment using a single-axis joint test rig.