Study of ROS-Based Autonomous Vehicles in Snow-Covered Roads by Means of Behavioral Cloning Using 3DCoAutoSim

Autonomous driving in winter weather conditions has always been a unique challenge, and as such it is an interesting research topic. Due to reasons related to safety and local laws, simulators have become one of the first choice for the required research. This paper extends the capabilities of the 3DCoAutoSim simulation platform with a realistic simulation environment for the study of autonomous driving with ROS-controlled vehicles in adverse weather conditions such as snow-covered roads. The weather-related details of the environment such as snow fall and car tracks on the snow were implemented by using Unity3D's physics and graphics engine. A series of autonomous driving experiments based on behavioral cloning were performed to test the performance of the environment and its scalability for ROS-based machine learning applications. Results from the experiments conducted to validate the approach demonstrated a good driving performance. Moreover, results from the model trained with the data set generated in the snowy environment, showed that car tracks features in the snow promoted the learning and generalization steps in the machine learning process.