HMSim: A Hierarchical Multi-Agent Learning-Based Simulator For Urban Driving Scenarios

Our framework has two hierarchical controllers, the high-level controller runs every K timesteps and produces goals for the low-level controller, which controls each agent per timestep. The high-level controller takes the scenario representation as input, including the map contexts and the agent history information. The high-level controller jointly infers the goal for each agents, and conditioned on that goal, the low-level controller generates actions to actuate the goal.

Our experiemnts show superior performance in simulation accuracy, diversity and plausibility.

We visualize our resimulated scenarios from waymo datasets.