Long-Term Interactive Driving Simulation: MPC to the Rescue

Simulation now plays an important role in the development of autonomous driving algorithms as it can significantly reduce the economical cost and ethical risk of real-world testing. However, building a high-quality driving simulator is not trivial as it calls for realistic interactive behaviors of road agents. Recently, several simulators employ interactive trajectory prediction models learnt in a data-driven manner. While they are successful in generating short-term interactive scenarios, the simulator quickly breaks down when the time horizon gets longer. We identify the reason behind: existing interactive trajectory predictors suffer from the out-of-domain (OOD) problem when recursively feeding predictions as the input back to the model. To this end, we propose to introduce a tailored model predictive control (MPC) module as a rescue into the state-of-the art interactive trajectory prediction model M2I, forming a new simulator named M \(^2\) Sim. Notably, M \(^2\) Sim can effectively address the OOD problem of long-term simulation by enforcing a flexible regularization that admits the replayed data, while still enjoying the diversity of data-driven predictions. We demonstrate the superiority of M \(^2\) Sim using both quantitative results and visualizations and release our data, code and models: https://github.com/0nhc/m2sim .