Physics-infused Intention Network for Crowd Simulation

Luping Li, Xin Guo, Yang Liu, Zhongming Jin, Xing Su, Jieping Ye

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: crowd simulation, physics-informed, agent-based pedestrian simulation

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: a novel physics-infused crowd simulation method.

Abstract: Crowd simulation has garnered significant attention in domains including traffic management, urban planning, and emergency management. Existing methods can be classified as either rule-based or learning-based approaches, with the former lacking authenticity and the latter lacking generalization. Recent research has attempted to combine these approaches and propose physics-infused methods to address the aforementioned limitations. However, they continue to adhere strictly to the framework of the physical model, neglecting to depict the attention mechanism as a critical component of behavior. This limitation results in deficiencies in both the fidelity and generalizability of the simulations. This paper introduces a novel framework called Physics-infused Intention NEtwork (PINE) for crowd simulation. Our model introduces a physical bias while endowing pedestrians with the ability to selectively enhance the fine-grained information most relevant to one’s current behavior. In addition, we design a variable-step rollout training approach with an optimized loss function to address cumulative errors in simulation. By conducting extensive experiments on four publicly available real-world datasets, we demonstrate that our PINE outperforms state-of-the-art simulation methods in accuracy, physical fidelity, and generalizability.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4676