Nonlinear Model Learning for Compensation and Feedforward Control of Real-World Hydraulic Actuators Using Gaussian Processes

Abstract

This paper presents a robust machine learning framework for modeling and control of hydraulic actuators. We identify several important challenges concerning learning accurate models of the dynamics for real machines, including noise and uncertainty in state measurements, nonlinear effects, input delays, and dataefficiency. In particular, we propose a dual-Gaussian process (GP) model architecture to learn a surrogate dynamics model of the actuator, and showcase the accuracy of predictions against the piecewise and neural network models that have been widely used in the literature. In addition, we provide robust techniques for learning neural network inverse models and controllers by batch GP inference in an automated, seamless and computationally fast manner. Finally, we demonstrate the performance of the trained controllers in real-world feedforward and tracking control applications.Index Terms-Hydraulic actuators, machine learning for robot control, model learning for control. I. INTRODUCTIONA LL real-world mechanical systems that provide a means to automation today are powered by integrated controllers. These controllers are carefully designed and fine-tuned to provide the best possible performance for the intended application, and their development follows a conventional process. First, the system dynamics are modelled, wherein prior knowledge and general understanding of the physics of the underlying process Manuscript

Multi-Log Grasping Using Reinforcement Learning and Virtual Visual Servoing

“…However, there are differences that make log grasping a special case, most notably regarding grasping multiple objects, the unstructured forest environment, the electro-hydraulic crane actuation, the system size, and exposure to the elements. For the specific application of log grasping and autonomous forwarding, there are good solutions for crane motion planning and control [9,10] without considering grapple-log interaction or surrounding obstacles. Reinforcement learning (RL) control has proven to be effective for the same task in simulations, grasping a single log with known pose [11].…”