How to simulate realistic, safety-critical traffic for AV testing?
Traffic Gamer: Reliable and Flexible Traffic Simulation for Safety-Critical Scenarios with Game-Theoretic Oracles
This paper introduces TrafficGamer, a novel algorithm for simulating realistic and diverse traffic scenarios, especially safety-critical ones involving multiple vehicles. It leverages game theory to model complex vehicle interactions and ensure fidelity to real-world data while adapting to different levels of competition and risk.
For LLM-based multi-agent systems, TrafficGamer offers a valuable tool for generating training data and testing scenarios. Its ability to simulate complex, interactive, and nuanced driving behaviors, going beyond simple imitation, can help train more robust and reliable LLMs for autonomous driving applications. Additionally, TrafficGamer's flexibility in controlling the intensity of competition and risk within scenarios allows for the creation of diverse, challenging test cases for evaluating the performance of multi-agent LLM systems in various situations.