- AI systems can finally model cause‑and‑effect when the set of objects and their links shift.
The paper defines relational structural causal models, an extension of Pearl’s classic framework that accounts for varying entities and relations. The authors prove that without extra assumptions, neither causal nor purely observational queries about unseen object groupings are identifiable. By introducing relational causal graphs and symbolic criteria, they show how identification becomes possible even with hidden confounders. A concrete implementation, relational neural causal models, is tested on synthetic traffic scenarios featuring differing numbers of cars, traffic lights, and pedestrians, where it beats non‑relational baselines.
This matters because most current causal AI assumes a fixed set of variables, limiting its use in dynamic environments like autonomous driving or robotics. The relational approach promises better generalization to novel configurations without retraining each time the scene changes.
Still, the work is confined to simulations; real‑world validation will reveal whether the symbolic criteria hold up amid noisy sensors and unpredictable human behavior.