Evaluation of Deep Reinforcement Learning Algorithms for Autonomous Driving

Once considered futuristic, machine learning is already integrated into our everyday life and will shape many areas of our daily life in the future: This success is mainly due to the progress in machine learning and the increase in computing power. While machine learning is used to solve partial problems in autonomous driving, the support of high-resolution maps severely limits the use of autonomous vehicles in unknown areas. At the same time, the structuring of the overall problem into modular subsystems for perception, self-localization, planning, and control limits the performance of the systems. A particularly promising alternative is end-to-end learning, which optimizes the system as a whole. In this work, we investigate the application of an end-to-end learning method for autonomous driving, employing reinforcement learning. For this purpose, a system is developed which allows the examination of different reinforcement learning approaches in a simulated environment. The system receives simulated images of the front camera as input and provides the control values for steering angle, accelerator, and brake pedal position as direct output. The desired behavior is learned automatically through interaction with the environment. The reward function is currently optimized for following a lane at the highest possible speed. Using specially modeled environments with different levels of detail, multiple deep reinforcement learning approaches are compared. Among other aspects, the extent to which a transferability of trained models to unknown environments is possible is examined. Our investigations show that Soft Actor-Critic is the best choice of the tested algorithms concerning learning speed and the ability to generalize to unseen environments.