Development of RL environment for large-scale autonomous driving tasks

This project aims to build a general RL training and testing environment for the autonomous driving tasks. My works: 1) Combined a sophisticated vehicle dynamics and the large-scale traffic and map constructed using SUMO to realize co-simulation of the ego vehicle and its surrounding environment. 2) Designed the basic RL elements in the field of autonomous driving following the interface defined in Gym, such as the state, the action, the reward function, and the reset conditions etc, which can be easily reused or modified depending on your task. 3) Correspondingly, an analytic model of the environment is developed to facilitate its use in model-based RL algorithms. 4) A cluster of training and testing environments for our own use are also included, for instance, environments with specific designed RL elements (such as using bird view image as state, or using trajectory as action), ones with different simulation parameters (such as vehicle parameters, or the map), as well as environments for multiple automated vehicles (such as the centralized control environment, and the distributed control environment).