Using the CARLA Simulator to Train A Deep Q Self-Driving Car to Control a Real-World Counterpart on A College Campus

Engineering notes

Key topics: self-driving car, self-driving, autonomous vehicle, sensor fusion, carla, level 4, radar, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous vehicles (AV) provide many benefits to society. They provide a way for people with the inability to drive a way to travel on their own. While there has been significant progress made in their field, AVs still struggle with handling difficult, urban environments. This study aims to implement a level 4 AV utilizing a deep q neural network within the Car Learning to Act (CARLA) simulator. The Carla simulator is used to train and test a self-driving car agent implemented with the Deep Q Neural Network (DQN). Sensor fusion is being implemented with three cameras located on the hood, left and right rearview mirrors, radars on the hood and trunk of the car, a GPS, and an Inertial Measurement Unit (IMU). Our simulated vehicle setup mimics a golf cart in the real world, equipped with the same sensor suite and setup. The vehicle is trained on Map 10 of Carla, with dynamic weather effects such as light, dark, clear skies, cloudy, dry, rainy, foggy, and windy. It is also trained with 5 other vehicles on the map as obstacles. Each simulation route is multipath and of random length and complexity. The agent maintains an average speed between 0 and 3.4 kph, on average makes it 7.64 meters away from its starting position, and has 119 meters left to reach the route destination.

Links and sources

• Official / arXiv page

Comments