Reproduce “Comprehensive Training and Evaluation on Deep Reinforcement Learning for Automated Driving in Various Simulated Driving Maneuvers” on OpenResearch

About this paper

Yongqi Dong, Tobias Datema, Vincent Wassenaar, Joris van de Weg, Cahit Tolga Kopar, Harim Suleman

Researchers at Delft University of Technology developed 'ComplexRoads', a novel integrated simulation environment, and custom reward functions to train Deep Reinforcement Learning (DRL) agents for automated driving. This approach resulted in models capable of generalizing across diverse driving maneuvers with improved safety and comfort compared to single-scenario training.

Key insights

• Customized reward functions, incorporating explicit cues for lane keeping and effective speed, consistently lead to superior DRL agent performance, particularly in safety metrics like on-lane rate.
• Training DRL agents on a diverse, integrated environment like 'ComplexRoads', especially with randomized starting locations, is effective in developing highly adaptable policies that generalize across multiple driving maneuvers.
• Trust Region Policy Optimization (TRPO) generally yields more robust and smoother driving policies compared to Deep Q-networks (DQN) when guided by an effectively designed reward function.

Summary by alphaXiv.