Deep Reinforcement Learning with external control: Self-driving car application

A Self-driving car using an end-to-end deep reinforcement learning[1] algorithms trained on lane-keeping task performs well in circuits that don't need decision making but cannot deal with situations like choosing to turn left or right in an upcoming crossroads, deciding when to leave a traffic circle or toward which path/destination to go. In this paper we propose a new Deep Reinforcement Learning architecture that supports external command as high-level input, that we call Steered Deep Reinforcement Learning (SDRL), we apply the SDRL architecture on the Deep Deterministic Policy Gradient algorithm DDPG and use CARLA a High-fidelity realistic driving simulator as a testbed environment to train and experiment the new model, since testing in ground truth turns out to be costly and risky. The Steered DDPG (SDDPG) model performs well on the road/roundabouts and responds correctly to the external commands that allow the driving agent to take the right turns.