Relational Structural Causal Models Advance Causal AI for Combinatorial Environments.

For artificial intelligence to truly understand its environment, it needs models that are both causal, allowing for reasoning about interventions and counterfactuals, and combinatorial, enabling generalization to novel arrangements of objects. This work formalizes and studies Relational Structural Causal Models (RSCMs), which extend the established Structural Causal Models to scenarios where objects and their relationships are dynamic and varied. The researchers demonstrate that identifying answers to both causal and observational queries about new object combinations is challenging without additional assumptions. To overcome this, they define relational causal graphs and derive symbolic criteria for identification, even in the presence of unobserved confounding variables. This provides a robust theoretical foundation for learning causal models in complex, object-centric environments. Finally, the paper proposes relational neural causal models, a practical approach that is proven correct and outperforms non-relational baselines. This was demonstrated in simulations of traffic scenes featuring diverse cars, signals, and pedestrians, highlighting the model's ability to generalize effectively to varying object configurations.