PathCluster: Pedestrian Group-Adaptive
Social Navigation in Dense Crowds

Nihal Gunukula · Aniket Bera

IROS 2025

TL;DR

Navigating a dense crowd one pedestrian at a time does not scale. PathCluster groups people with similar trajectories and plans against the groups instead of the individuals: a 45 percent higher success rate, a 25 percent lower collision rate, and harder navigation tasks solved within a 48 hour compute budget than prior social navigation models in extremely crowded environments.

The idea

Humans do not navigate crowds by dodging individuals. We read structure: the group of three walking together that should not be split, the couple stopped to look at something, the stream of commuters you merge into rather than cut across. PathCluster gives a robot the same primitive. A group generator identifies clusters of pedestrians with similar trajectories in real time and treats each cluster as a cohesive unit, which collapses trajectory prediction from hundreds of individuals to a handful of groups while keeping the benefit of prediction.

Results

• 45 percent higher success rate than prior social navigation models in extremely crowded environments.
• 25 percent lower collision rate under the same conditions.
• Harder tasks become tractable: navigation problems that prior methods could not solve within a 48 hour limit are completed by PathCluster.
• The learned policy picks up the judgment people never articulate: when to yield, when to merge, when to wait.

BibTeX

@inproceedings{gunukula2025pathcluster,
  author    = {Gunukula, Nihal and Bera, Aniket},
  title     = {PathCluster: Pedestrian Group-Adaptive Social Navigation in Dense Crowds},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2025}
}

Work with the IDEAS Lab at Purdue University.