Artificial Intelligence Controls for Vehicle Emergency Maneuvering

Autonomous driving is the future of automotive. The most important challenge in autonomous driving applications is represented by safety. An autonomous driving system to be perceived safe has to be trustful in emergency situations. To recognize an emergency situation and act immediately to avoid collisions is one of the most important features of an autonomous driving system. This thesis work investigated the topic of collision avoidance in emergency situation in autonomous driving using deep reinforcement learning. According to the literature related to this problem, two kinds of ADAS controls were studied. The studied controls were: Autonomous Emergency Steering and Autonomous Emergency Braking. The use of reinforcement learning based methods in autonomous driving field is increasing, because of the versatility of these algorithms. The deep reinforcement learning algorithm used in this thesis work was a Double Deep Q-Network with Dueling Architecture. This model-free deep reinforcement learning algorithm allowed the reinforcement learning agent to learn complex controls in realistic simulated environment. To provide the most realistic environment possible, a simulator with state of the art rendering quality was used. The problem was contextualized in an urban scenario where a pedestrian crossed the street at an unexpected timing. The information were provided to the agent through sensors. Cameras, collision and lane invasion sensors were used. The information used by the agent to learn the policy were provided by the camera sensor. The output frames of the camera were processed and stacked together to give the agent the sense of motion. The outputs of the collision and the lane invasion sensors were used to determine the terminal state of the learning episodes. The settings of the tests and their main parameters were chosen taking as reference the Euro NCAP AEB test protocol. The agent successfully learned from scratch how to perform an emergency brake action and an emergency overtake action. In the Autonomous Emergency Steering test the agent also learned from scratch how to keep the lane. This thesis work makes contribution on the topic of investigating collision avoidance in emergency situation using deep reinforcement learning. Since the input was given to the agent through a low cost and established sensor, the trained agent can be used in real vehicles for further studies.