Predicting one movement and reproducing a flock are different jobs. Across six noise-free simulations of 30 or 50 moving particles, PR-GNN recorded the lowest aggregate position error over 500 steps—even though rivals sometimes predicted the next step better.[1]
The mechanism has a long history in computer graphics. Craig Reynolds created his Boids model in 1986 around three local behaviours: avoiding crowded neighbours, aligning with nearby movement and moving towards nearby flockmates. A coherent flock emerged from those individual responses. His account records modified Boids software producing bat swarms and marching penguins for Batman Returns.[2]
That history explains the attraction of local rules. An animator need not draw a complete route for every animal. Specify how each responds to its neighbours, and a group pattern can develop. Reynolds also describes a goal-seeking behaviour for directing a flock along a scripted path. The overall scene and the local motion can therefore be controlled at different levels.[2]
Learning the interactions
The new study, released as an accepted journal article on 26 October, learns from trajectories. Its network gives different neighbours different weights, then separates acceleration into self-motion, velocity alignment and spatial interaction. Training follows several predicted steps and penalises changes in the group’s alignment and spread.[1]
An illustrative comparison is copying a route one instruction at a time. If the next instruction always starts from the correct position, a small mistake can disappear at the next check. If it starts from where the traveller actually ended up, mistakes alter every later instruction. A long simulation makes the second situation unavoidable.
PR-GNN reproduced both coordinated travel and circling patterns in the tested systems. The experiments used synthetic trajectories, not tracked animals. Learned attention highlighted relevant neighbours without exactly recovering the original interaction rules. Those are distinct achievements: reproducing a group and identifying the precise rule that generated it.[1]
Training on the group’s organisation as well as individual positions kept the simulated motion coherent over the full test.[1]
Sources & further reading
- Learning collective motion dynamics from trajectory data with physics-regularized graph neural networks Complex & Intelligent Systems · 26 Oct 2026 · Accessed 28 Oct 2026
- Boids: Background and Update Craig Reynolds · Accessed 28 Oct 2026