Learning feed-forward control for a Soft Robotics phantom to simulate respiratory liver motion

Finished: 2019-07-04

Individual assignment

Liver cancer, also known as hepatic cancer is one of the leading factors of cancer deaths, having an annual count of 782 000 deaths worldwide. At earlier stages, treatments include partial hepatectomy or transplantation. An increasing amount of robotic assistance systems are being a developed in order to improve the outcome of these interventions, these benefits include smaller scars and potential faster recovery times. In order to provide a test platform for such robotic assistance systems, a former project at the University of Twente focused on developed a liver phantom which is able to simulate respiratory motion of the liver. Since the liver is placed directly under the diaphragm, respiratory motion induces considerable movement of the liver. This makes percutaneous treatment complicated. The platform earlier described can help test the ability of compensating for respiratory motion for these robotic assistance systems, but also for surgeons in training. Besides that, the phantom could also be used for an accurate substitution of a real human in order to test and develop for example Magnetic Resonance Imaging (MRI) techniques.

The current phantom setup can mimic the motion of respiratory behaviour by using a feed-forward controller for a single situation. Changing any of the setup parameters, or even reapplying the substitute skin can result in a different trajectory of the liver and subsequently the tumor. This makes the repeatability of an experiment low, even tough this is crucial for good test results. This problem arises because the feed-forward controller is not based on a model, but tuned on a "black-box" principle meaning that the desired output is achieved by tuning input parameters manually. There is nothing wrong with this principle, however if parameters are changed there is really no way of achieving the desired trajectory rather than starting the tuning process over again. This process is time consuming and since the process is manual the precision is lower.

The goal of this project is to achieve greater repeatability and performance by introducing self learning feed-forward controller, in which the effect of the system parameters will be characterized. In this way a change in setup parameters can be easily accounted for in the controller. Besides that, different desired trajectories can be achieved and real human motion measurements can be mimicked by the proposed controller. Overall the system will have far improved repeatability and greater precision.