DRIVING STRATEGY FOR AUTONOMOUS VEHICLE IN DEALING WITH EMERGENCY SCENARIOS BASED ON DEEP REINFORCEMENT LEARNING

• Main Author: Mardhatillah, Irina
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/77498
• Institution: Institut Teknologi Bandung

Autonomous vehicles have not been fully reliable in dealing with traffic uncertainties until recently. Several recent studies have proposed several systems to improve safety in autonomous vehicles, but these studies only focus on autonomous vehicles and do not specifically analyze emergency scenarios that may occur. Thus, this final project research focuses on developing driving strategies for autonomous vehicles in dealing with emergency scenarios at intersections. Autonomous vehicle is trained using the proximal policy optimization (PPO) algorithm in a simulation environment built using the CARLA Simulator. The vehicle is trained to avoid collisions or at least minimize the impact of collisions when dealing with pre-crash scenarios at intersections, as defined by the National Highway Traffic Safety Administration (NHTSA) in its report entitled "Pre-Crash Scenario Typology for Crash Avoidance Research". The simulation in this study includes training and testing phases carried out at two types of intersections, i.e., signalized intersections and unsignalized intersections, as well as three levels of traffic complexity, i.e., low, medium, and high, that are differentiated based on the traffic components. Autonomous vehicle is trained to control throttle, brake, and steering based on information regarding the surrounding environment obtained from sensors integrated into the vehicle. The simulation results show that the overall performance of autonomous vehicles provides the highest success rate in avoiding collisions at 85%. The ability of autonomous vehicles to avoid collisions decreases as the traffic complexity increases. Collisions still occur due to a side of the vehicle that is not captured by the camera, so the autonomous vehicle cannot detect vehicles moving on that side. However, the vehicle tend to move at lower speeds when dealing with more complex traffic, which is indicated by a smaller average value of impulse due to collisions when autonomous vehicle is in more complex traffic.