A reinforcement learning model for autonomous vehicles with realistic car simulation in urban using Unity

Abstract Simulator training with reinforcement learning (RL) for autonomous vehicles (AVs) offers advantages over supervised learning. However, transferring the learned behaviours to the real world is a challenging task due to the inconsistencies between the data captured by the vehicle’s sensors in the simulated environment and the real world. Additionally, some of the sensors that the AVs rely on are sensitive to weather and lighting conditions. Our proposed model addresses the challenges of sensor data disparity and environmental variation. It utilizes three sensing components which are radio detection and ranging (RADAR), inertial measurement units (IMUs), and global positioning systems (GPSs) to overcome the addressed drawbacks. The proposed model incorporates a carefully designed reward system and prioritizes computational efficiency by using fewer number of sensors and ensures safe and efficient driving. The chosen sensors enable easier knowledge transfer from the simulator to the real-world due to their consistent data representation. The model leverages the Unity engine and ML agent to train AVs for both urban and highway environments. The experimental results show that our suggested model effectively trained AVs to navigate through complex urban areas without collisions while keeping them in their lanes. The demonstration video is provided in the following link: https://youtu.be/YCOjli7lrCM