Query-Conditioned World Models for Embodied AI

A recent paper on arXiv contends that world models for embodied AI should be physically plausible, aimed at addressing intervention queries by accurately depicting the physical framework that influences action results, rather than simply forecasting future observations. The authors highlight a fundamental flaw in current observation-predictive models: different physical systems may appear the same but can behave differently when intervened upon, resulting in visually convincing yet physically incorrect predictions. Benchmarks that maintain a consistent visible scene while altering latent physics reveal that these models might suggest impractical actions, miscalculate interaction results, or endorse unsafe behaviors. The paper advocates for world models in embodied AI that pinpoint the most straightforward physical abstraction necessary for intervention queries, incorporating modular elements such as environment representation, latent state and parameter estimation, action specification, and interventional reasoning.