UC Davis

In this research, we develop a framework for autonomous vehicles to interact safely with pedestrians in urban scenarios based on a game theoretic approach. The primary idea is to simplify the complex interaction and capture distinct behavioral features of pedestrians. The game setup incorporates a feedback control system with nonlinear dynamics. The proposed concept works to linearize the nonlinear part of the player dynamics in each iteration and uses the quadratic cost to formulate the interaction patterns at the urban streets. The algorithm takes inspiration from the iterative linear quadratic form and builds on the stochastic game to find Nash Equilibrium. We divide the dissertation research into three approaches. In the first approach, we derive the Algebraic Riccati equation for the stochastic non-zero-sum (NZS) game to capture the competitive and cooperative nature of the interaction between players. We explore distinct features of pedestrians and develop the interaction framework for AV to capture these behavioral attributes, including conservative, aggressive, and grouping. After tackling the noise in player dynamics in the first approach, we investigate the external noise during an interaction event and define a robust game framework. The game setup allows for a finite disturbance in the player cost function and noise in the system dynamics. The disturbance in the cost function tries to maximize the overall cost as an adversarial input, whereas the noise perturbs the system dynamics. In the final approach, we extend the interaction framework with a finite delay in the feedback control of the players to reflect real-world scenarios. The interaction game framework developed in this research could complement the decision-making system of autonomous driving in urban streets to ensure accurate motion planning and safe interaction with respect to pedestrians.