Data augmentation to enhance human-robot interaction

Abstract

Contemporary research in robotics is focusing on autonomous and independent robots that can operate in unconstrained environments and interact with humans. Perceiving of and acting within these environments not only creates novel challenges but provides new opportunities to improve state-of-the-art of robotics. This thesis presents one such opportunity by utilising information from Human-Robot Interaction for data augmentation. Data augmentation is a valuable tool to improve recognition approaches, especially in complex environments. Recent research in machine learning is often driven by the mantra "we need more data". In contrast, data augmentation can outperform state of-the-art methods provided additional information can be used without the need to acquire more input data. In Human-Robot Interaction, we benefit from knowing the context in which a robot acts and perceives information. The presented work takes a closer look at such use-cases, introduces the notion of direct and indirect data augmentation, and highlights the importance of data augmentation by providing experiments on datasets that were created from interactions with the iCub humanoid robot. This thesis shows the application of data augmentation and provides improvements for state-of-the-art action and object recognition in the context of Human-Robot Interaction.