Kei Yoshida¹ (), Sina Feldmann¹, William H. Warren¹; ¹Brown University

When following a crowd, pedestrians combine information from multiple neighbors in their visual neighborhood to guide walking, consistent with distance-weighted averaging (the SCruM model of Self-organized Collective Motion, Rio et al., 2018). Last year, we reported a real-crowd experiment in which four confederates attempted to steer or split a group of walking participants (N=16–22). When participants were unaware of the confederates (covert), they could steer but not split the crowd, consistent with weighted averaging. But when participants were instructed to follow confederates carrying flags (explicit), they generated both steering and splitting. To account for explicit leaders, here we extend the weighted-averaging account by incorporating attention-based weights into the SCruM model. Modified models were tested using multiagent simulations, which took confederate trajectories as input and computed simulated participant trajectories. We then compared the simulated final headings with the human data. For covert confederates, the original SCruM model reproduced the empirical patterns of steering (KS test, p < .001 for left and right) and not splitting (overlapping GMM components with means -0.86° and 0.24°). However, the model could not replicate the strong influence of explicit confederates. Forcing the confederate weights to 1 reproduced the large left and right turns observed with explicit confederates (p < .001 for left and right). But splitting only emerged when agents allocated their full attention to the nearest confederate and the weights of the other three confederates were set to zero (two GMM components with means -8.0° and 8.4°). Neighbors retained their distance-based weights, yielding coherent sub-group motion. These findings support distance-weighted averaging as a generally robust mechanism for collective motion, but the addition of attention-based weights is needed to explain collective behavior in scenarios with explicit leaders. The results imply that followers attend to a single nearest leader in a crowd.

Acknowledgements: Supported by NSF BCS-1849446, NIH 1S10OD025181