Exploiting Structure in Learning: A Path Toward Building Safe and Adaptive Robots

Abstract

As robots venture into real-world applications, there is an increasing need for them to effectively learn from experience and adapt to unseen situations. This thesis addresses a few critical challenges to practical robot learning, including safety and sample efficiency. The main perspective of this thesis is to leverage various types of structure, ranging from explicit domain knowledge about robotics problem, such as system dynamics, task decomposition, etc., to hidden structure in robotics data. Such structure can provide valuable insight into solving robotics tasks, but cannot be directly used by most off-the-shelf learning-based solutions. The first part of this thesis focuses on encoding known domain knowledge into structured policy classes for learning, giving standard off-the-shelf learning algorithms the ability to admit formal safety guarantees, learn efficiently with small datasets, and generalize well to new conditions. The second part of this thesis considers the structure in robotics data. This thesis considers two types of structure in data: 1) explicit structure given by what information each subset of data provides and 2) implicit structure induced by common data collection processes. By reasoning about the explicit structure in data, this thesis introduces a general learning paradigm and an associated learning algorithm which have theoretical guarantees and work well empirically. Finally, this thesis shows that implicit structure from data collection processes can be sometimes unintentionally leveraged by learning algorithms to achieve seemingly surprising robustness and safety properties.