Multi-Agent Imitation Learning for Driving Simulation

摘要: Simulation is an appealing option for validating the safety of autonomous vehicles. Generative Adversarial Imitation Learning (GAIL) has recently been shown to learn representative human driver models. These models were learned through training in single-agent environments, but they have difficulty generalizing multi-agent driving scenarios. We argue these difficulties arise because observations at and test time are sampled from different distributions. This difference makes such unsuitable simulation scenes, where multiple agents must interact realistically over long horizons. extend GAIL address shortcomings a parameter-sharing approach grounded curriculum learning. Compared with policies, policies generated by our PS-GAIL method prove superior interacting stably setting capturing emergent behavior drivers.