Hybrid-State System Modelling for Control, Estimation and Prediction in Vehicular Autonomy

This thesis studies the Hybrid-State System models and their properties for different pieces of the urban autonomy problem. For autonomous vehicles that operate in real-life, mixed-mode traffic, a number of parallels between the human-driven system and the autonomous counterpart can be identified and captured in the hybrid-state system setting. For the control subproblem of the urban autonomy, this thesis proposes a system architecture, related design approaches for autonomous mobile systems and guidelines for self-sufficient operation. Development of a tiered layout for a hybrid-state control in a series of stages as well as the integration of such a controller in the overall autonomy structure are proposed and demonstrated as part of multiple examples, including The Ohio State University participation in Defense Advanced Research Projects Agency Urban Challenge 2007. The hierarchical layout and the iterative design methodology enable design flexibility through compartmentalization of the overall system and helps prepare for various contingencies, as illustrated on specific development cycles. The sensing and perception part of the autonomy implementation relies on a probabilistic hybrid-state system modelling method that is developed for driver-behavior analysis and prediction. The model fits into and captures the central modules of the existing Human Driver Model. The stochastic models, based on the observable actions of the driver/vehicle interaction, are useful in representing the behavior of human-driven vehicles in certain urban decision-making ii scenarios. The Driver Intention Estimator presented utilizes the developed stochastic models to detect and predict high-level, abstract decisions of observed drivers through traffic scenarios and it can be expanded to form scenario safety estimation tools as demonstrated. As for the analysis of the developed estimators and as a useful tool for hybrid-state systems in general, this study develops an encoding scheme for discretestate systems as part of a hybrid-state hierarchy. The codes are command-based, in the sense that the interactions of the discrete states with the continuous states are exploited to attach significance to what each discrete state does to the continuous subsystem. The resultant codeset is independent of how the discrete-state transitions are designed and conventional binary tools such as truth tables and K-maps are easily applicable in the binary representation of the codes. Code-based representation of every possible combination of commands/behaviors governed by the discrete subsystem is useful in a number of design scenarios, an example of which is the generation of a consistent norm for discrete states. Such a norm is demonstrated to be useful in hybrid-state estimation.