Real-World Navigation for Intelligent AR Characters

Abstract

As Augmented Reality (AR) technologies are maturing and AR capable consumer devices are readily available, digital content embedded into the physical world can be made accessible to many people. Nonetheless, intelligent virtual characters that help us with everyday tasks are still largely unexplored. To fill this gap, the GTC lab at ETH Zurich is developing solutions to enable rich AR experiences using intelligent agents. One important skill of intelligent characters is navigating the physical environment. To investigate this feature, this thesis develops an AR prototype for guiding people.

Our approach improves on the work of Thomas Lang [Lang, 2021] who already implemented an early AR prototype for user guidance. We expand on this by focusing on three issues under- lying the actual navigation: establishing common points of the real world between different AR sessions, persisting the physical environment for navigation, and aligning content from foreign coordinate systems.

For establishing common points, we add a more universal technology that does not depend on having images in the real world, as the early prototype does. For navigation, we move on from an explicit graph that requires a cumbersome manual creation, to automatically computing the walkable area based on the physical environment. For aligning content, we build a more accurate and robust solution that uses multiple points of reference instead of just a single one.

To evaluate how useful our prototype is and how well it works, we conduct a user study. For this, multiple people test the prototype in different locations. Through usage data logged in the background and by getting feedback from the participants directly, we examine the strengths and weaknesses of the system. Despite running into imperfections that hindered some users from completing their targets, the AR experience was overwhelmingly positively received by the participants.