Are socially-aware trajectory prediction models really socially-aware?

Our transportation field has recently witnessed an arms race of neural network-based trajectory predictors. While these predictors are at the core many applications such as autonomous navigation or pedestrian flow simulations, their adversarial robustness not been carefully studied. In this paper, we introduce a socially-attended attack to assess social understanding prediction models in terms collision avoidance. An is small yet carefully-crafted perturbations fail Technically, define failure mode output, and propose hard- soft-attention mechanisms guide our attack. Thanks attack, shed light on limitations current understanding. We demonstrate strengths method recent models. Finally, show that can be employed increase state-of-the-art The code available https://s-attack.github.io/.