Codebase for our Hybrid Deep Reinforcement Learning (H-DRL) based automated driving project. The related paper can be accessed with this link.
An overview of our framework. The proposed system is a hybrid of a model-based planner and a model-free DRL agent. *Other sensor inputs can be anything the conventional pipe needs. ** We integrate model-based planners into the DRL agent by adding "distance to the closest waypoint" to our state-space, where the path planner gives the closest waypoint. Furthermore, the reward function is modified accordingly: the agent is penalized for straying away from the model-based planners' waypoints and also making a collision. Any kind of path planner can be integrated into the DRL agent with the proposed method.
This project was forked from a conventional DRL implementation for CARLA by Sentdex. https://github.com/Sentdex/Carla-RL