Autonomous vehicle control in CARLA Challenge

Abstract

The introduction of Autonomous Vehicles (AVs) in a realistic urban environment is an ambitious objective. AV validation on real scenarios involving actual objects such as cars or pedestrians in a wide range of traffic cases would escalate the cost and could generate hazardous situations. Consequently, autonomous driving simulators are quickly evolving to cover the gap to achieve a fully autonomous driving architecture validation. Most used 3D simulators in self-driving cars field are V-REP (Rohmer, E., 2013) and Gazebo (KOENIG, N. and HOWARD, A., 2004), due to an easy integration with ROS (QUIGLEY, 2009) platform to increase the interoperability with other systems. Those simulators provide accurate motion information (more appropriate for easier scenes like robotic arms) but not a realistic appearance and not allowing real-time systems, not being able to recreate complex traffic scenes. CARLA (DOSOVITSKIY, A., 2017) opensource AV simulator is designed to be able to train and validate control and perception algorithms in complex traffic scenarios with hyper-realistic environments. CARLA simulator allows to easily modify on-board sensors such as cameras or LiDAR, weather conditions and also the traffic scene to perform specific traffic cases. In Summer 2019, CARLA launched its driving challenge to allow everyone to test their own control techniques under the same traffic scenarios, scoring its performance regarding traffic rules. In this paper, the Robesafe researching group approach will be explained, detailing vehicle motion control and object detection adapted from Smart Elderly Car (GÓMEZ-HUÉLAMO, C., 2019) that lead the group to reach the 4th place in Track 3 challenge, where HD Map, Waypoints and environmental sensors data (LiDAR, RGB cameras and GPS) were provided.