Deep Reinforcement Learning Control Strategy at Roundabout for i-CAR Autonomous Car

Abstract

Institut Teknologi Sepuluh Nopember (ITS) has successfully implemented an Autonomous Car (i-CAR) as a mode of commuter transportation within the campus. One of the difficulties in implementing the maneuver is when passing roundabouts which are scattered on the i-CAR route. This paper discusses i-CAR’s maneuvers on roundabouts by implementing Deep Reinforcement Learning. The i-CAR vehicle is modeled in the CARLA simulation environment, then tested in a virtual environment in roundabouts with intersections and roundabouts without intersections that simulate U-turn maneuvers. The Deep Reinforcement method used is Deep Que Network (DQN) using various reward function configurations. Through experiments using CARLA simulations, it was obtained that Icar was able to pass roundabouts with and without intersections with an average deviation angle of 27.011 degrees and 30.068 degrees, respectively. The average time needed to pass the roundabout is 13.3 seconds and 7.9 seconds, with an average speed of 27.0 kmph and 28.5 kmph. This speed is still acceptable on campus, where the driving speed inside is limited to 40 kmph.