| Guidance behavior generated either by artificial agents or humans has been actively studied in the fields of both robotics and cognitive science. The goals of these two fields are different. The former is the automatic generation of appropriate or even optimal behavior, while the latter is the understanding of the underlying mechanism. Their challenges, though, are closely related, the most important one being the lack of a unified, formal and grounded framework where the guidance behavior can be modeled and studied.;This dissertation presents such a framework. In this framework, guidance behavior is analyzed as the closed-loop dynamics of the whole agent-environment system. The resulting dynamics give rise to interaction patterns. The central points of this dissertation are that: first of all, these patterns, which can be explained in terms of symmetries that are inherent to the guidance behavior, provide building blocks for the organization of behavior; second, the existence of these patterns and humans' organization of their guidance behavior based on these patterns are the reasons that humans can generate successful behavior in spite of all the complexities involved in the planning and control.;This dissertation first gives an overview of the challenges existing in both scientific endeavors, such as human and animal spatial behavior study, and engineering endeavors, such as autonomous guidance system design. It then lays out the foundation for our formal framework, which states that guidance behavior should be interpreted as the collection of the closed-loop dynamics resulting from the agent's interaction with the environment. The following, illustrated by examples of three different UAVs, shows that the study of the closed-loop dynamics should not be done without the consideration of vehicle dynamics, as is the common practice in some of the studies in both autonomous guidance and human behavior analysis. The framework, the core concepts of which are symmetries and interaction patterns, is then elaborated on with the example of Dubins' vehicle's guidance behavior. The dissertation then describes the details of the agile human guidance experiments using miniature helicopters, the technique that is developed for the analysis of the experimental data and the analysis results. The results confirm that human guidance behavior indeed exhibits invariance as defined by interaction patterns. Subsequently, the behavior in each interaction pattern is investigated using piecewise affine model identification. Combined, the results provide a natural and formal decomposition of the behavior that can be unified under a hierarchical hidden Markov model.;By employing the languages of dynamical system and control and by adopting algorithms from system identification and machine learning, the framework presented in this dissertation provides a fertile ground where these different disciplines can meet. It also promises multiple potential directions where future research can be headed. |
