Creating Behavior Trees for Autonomous Versatile Robots

This thesis tackles the long-standing problem of writing computer programs to control robots, ideally generating programs automatically without actually programming at all. For motivation, we notice that in recent years, two strong trends in industry are those of collaborative and mobile robotics. These application types share a critical property where robot programs now have to autonomously react to diverse events. Autonomous robots that operate without regular human intervention or guidance must therefore be controlled by policies that execute different actions depending on the circumstances, but creating a robust policy is considerably more difficult than writing a linear program that always executes the same instructions. There is also an established global industrial trend of more flexible production and smaller batches, leading to programming and commissioning costs becoming a larger share of the total cost of a robot application. All of this put together has increased the already high pressure to make robots easier and faster to program, or in other words, to make them more versatile. A popular and growing policy architecture used in robotics is behavior trees, driven by their inherent reactivity, transparency, and modularity. How to create behavior trees easily, however, is a highly active research topic, with many methods proposed from varying fields such as machine learning, automated planning, and intuitive user interfaces. In this thesis we study how we can improve and combine various methods such that they complement each other and the resulting system performs better. We propose several composite systems and validate their effectiveness in creating behavior tree policies in multiple robotic benchmark experiments.