A driving profile recommender system for autonomous driving using sensor data and reinforcement learning

The design of algorithms for autonomous vehicles includes a wide range of machine learning tasks including scene perception by the visual input from cameras and other sensors, monitoring and prediction of the driver and passengers’ state, and others. The aim of the present work is to study the task of personalizing the driving experience in an autonomous vehicle, taking into account the particularities and differences of each person in how he/she perceives the vehicle’s velocity. For this purpose, we employ the Actor-Critic Reinforcement Learning technique in order to automatically select the best driving mode during driving. The input to the actor-critic model comprises the driver’s stress and excitement, which are affected by the route conditions, and the vehicle velocity and angular velocity. The output at each step is the best mode for each driver, which better balances stress, excitement, and route completion time. The whole setup is simulated and tested within the Carla open-source simulator for autonomous driving research.