Crowd simulation, the study of the move of crowded agents in various environments, presents an application domain for machine learning. One challenge is to interpolate trajectories of multiple agents to approximate true trajectories while preserving collision-free property. Previous methods jointly optimize coupled trajectories and are computationally expensive. In this study, a new approach is proposed to decouple trajectories by encoding spatial dependencies among trajectories into a data-driven prior velocity. Simulated experiments on various combinations of priors and optimizers show speed up and the importance of the global flow prior.

  Another challenge in crowd simulation is to imitate the move of crowded expert agents. Previously it was not clear how data and training methods affect models when generalized to substantially different scenarios. In this work, we study behavior cloning (BC) method and a more sophisticated Generative Adversarial Imitation Learning (GAIL) method, on three typical types of scenarios. Simulated results suggest that (i) simpler training methods are better than more complex training methods, (ii) expert samples with diverse agent-obstacle and agent-agent interactions are beneficial for reducing collisions when the trained models are applied to new scenarios.