Interactive Pedestrian Simulation in iGibson

We present a learning-based local controller for pedestrian simulation in iGibson. We explore the feasibility of leveraging datasets of pedestrian trajectories to learn a model for socially-aware pedestrian simulation in indoor constrained environments. Based on previous work on pedestrian prediction, we augment Social-GAN with information on static obstacles, pedestrian trajectory histories, and pedestrian goals in order to use the architecture for human-like trajectory generation. In ongoing work, we empirically find that his local controller exhibits more realistic interactive behaviors than commonly used models in social navigation research such as ORCA.