Decision Making Framework for Autonomous Vehicle Navigation

A key trait of an autonomous vehicle is the ability to handle multiple objectives, such as planning to guide the autonomous vehicle from an initial to a goal configuration, avoiding obstacles, and decision making to choose an optimal action policy. Moreover, the autonomous vehicle should act in ways that are robust to sorts of uncertainties, such as wheel slip, sensors affected by noise, obstacle move unpredictably, etc. In this paper we propose a decision making framework for autonomous vehicle conducting a nonholonomic motion planner, localization, obstacle avoidance, and also dealing with the uncertainties. The decision making framework manages the safety and task related assignment by adopting the partially observable Markov decision process model. We predict N-steps belief states from the action sequence candidate and use them as inputs of a proposed reward value function. The dimensionality problem in the searching of an action sequence from the action space (linear and rotational velocities) is simplified by applying a nonholonomic motion planner as a reference. The simulation results show the decision path resulting from decision making framework depends on the initial setting of belief state of its position and orientation, and also the determination of uncertainties of the sensors and the actuators of the vehicle.