Human-like control for offshore excavators

Abstract

Offshore excavators are large hydraulically driven machines which are difficult to control due to slow dynamics, inherent nonlinearities, a varying environment and complex kinematics. As digging is performed under water, only limited visual feedback of the task can be provided by means of a visualization interface. Operators require an extensive amount of practice before being capable of achieving sufficient and consistent performance. Often, automation is implemented as a means of reducing costs related to expensive operators and attaining consistent performance. However, automation struggles with adapting to unforeseen situations and a large task variety, which are areas human operators excel in. Instead of attempting to fully automate excavators, this thesis takes a more human-centered approach, and focuses on the design and evaluation of a human-like controller to partially automate excavator operations, while assuming a human operator is still present to trade or share control with. In order to simultaneously deal with the various nonlinearities in the system while providing human-like control this work proposes the use of an Adaptive Model Predictive Controller, whose underlying principles are similar to those of humans.To determine whether the controller is indeed human-like a complex excavator model including a realistic soil model was developed and used to implement and tune the controller. Finally, a simulator experiment was conducted to compare the subjects and the controller in terms of performance for various tasks and the control behavior similarity for a well-trained task. Eight subjects controlled the excavator model and performed four stages, starting with a familiarization stage in which the subject got accustomed to the system. The other three stages (easy, difficult, boulder) featured a 9 m long target path, with conditions of varying difficulty between stages. The controller showed 2 to 3 times lower tracking errors for both the easy and difficult stage while providing 1.5 to 5 times smoother inputs, but could not overcome the unforeseen boulder whereas all subjects could, showcasing the importance of having humans and automation complement each other. Furthermore, a high quality fit (VAF > 70%) was found between the boom inputs of the subjects and the controller in the well-trained easy stage, indicating human-like control.