Using V2X and reinforcement learning to improve autonomous vehicles algorithms with CARLA
Type of DegreeMaster's Thesis
Computer Science and Software Engineering
MetadataShow full item record
Autonomous vehicles (AV) or cars of the future are only growing in popularity. However, there is a reported lack of trust in these AV. According to a recent survey conducted by the AAA automotive group on understanding people’s attitudes towards self-driving cars they found out that only 14% of drivers would trust and feel safe riding in an autonomous vehicle. Current autonomous vehicles rely on sensors such as RGB cameras, LiDAR, RADAR, and more. These sensors have limited perception and prediction capabilities in certain ambient conditions. This research aims to study the impact of connecting self-driving cars with their surrounding through Vehicle-to-Everything (V2X) data. V2X is a communication system where data from sensors, traffic lights and many other sources travel through a high-bandwidth and low latency network and can be used as input for autonomous cars. We expand on this by introducing a reinforcement learning (RL) algorithm that benefits from the use of V2X and trains a car in a simulated testbed on various maneuvering scenarios to emphasize the impact of using V2X compared to the use of traditional sensors alone. Furthermore, we compare our solution with a simple algorithm that relies on the use of a camera (RGB) sensor in various lighting and weather conditions. We use the open-source simulation software CARLA to validate the improvement of the autonomous vehicle algorithm coupled with V2X and RL.