The paper presents a method called Reinforcing Traffic Rules (RTR) for simulating realistic traffic scenarios, which is essential for the development of self-driving software. Traditional imitation learning methods often result in unrealistic traffic infractions, while reinforcement learning results in unhuman-like driving. RTR combines these approaches, learning from both real-world scenarios and procedurally generated long-tail scenarios to produce more realistic and generalizable traffic simulations. The study finds that the RTR method significantly improves the tradeoffs between human-like driving and traffic compliance.
Publication date: 2 Nov 2023
Project Page: https://waabi.ai/rtr
Paper: https://arxiv.org/pdf/2311.01394
