Visual-Servoing based Mixture of Experts Fixture Learning

Abstract

Haptic guidance plays an important role in the interaction between the user and the robot. It can be implemented in either teleoperation or hands-on devices. In both cases, haptic guidance assists the user in completing the task. Different types of guidance are required depending on the nature of the task. These are called Virtual Fixtures. For example, one Virtual Fixture often provides path guidance to lead the user to the workplace or between different workplaces. Also, forbidden regions for the robot could be defined. Vision-based guidance, such as Visual Servoing Fixtures, is necessary to adapt to environmental changes. The Visual Servoing Fixtures used in this work can even update the Virtual Fixture live. This Visual Servoing Fixture is based on a Mixture of Experts (MoE). This approach proposed by Mühlbauer et al. [1] mediates between several targets with different uncertainties. One problem with the Visual Servoing Fixtures is that several hyperparameters must be tuned for every use case. The tuning of these hyperparameters is a challenging task because the parameters influence each other, and the user needs to know the mathematics behind them. The goal of this Master’s Thesis is to estimate the hyperparameters for the Visual Servoing Fixtures using Learning from Demonstration (LfD). Finally to make adapting the Visual Servoing Fixtures to other tasks easier.