Physically Viable World Models: A Case for Query-Conditioned Embodied AI

2026-05-28

AI Summary

Existing embodied AI world models typically predict observations but struggle to correctly answer interventional queries, often producing physically infeasible behaviors when the underlying physical structure changes. This work proposes a physically feasible world model explicitly designed for interventional reasoning, grounded in a query-centric principle of physical abstraction: it employs the minimal yet sufficient physical representation needed to distinguish outcomes under interventions. The architecture is modular—comprising environment representation, latent state estimation, and interventional dynamics—and features a dynamic orchestration mechanism that adaptively composes analytical, simulation-based, and learned components as required by the query. The resulting model accurately answers interventional questions where existing systems fail, yielding physically consistent and auditable behavior recommendations. This approach establishes a new paradigm for planning, control, and safety verification, while introducing a feasibility criterion for world model design.

Abstract

World models for embodied AI must be physically viable: constructed to answer intervention queries by representing the physical structure governing action outcomes, rather than merely predicting future observations. Existing observation-predictive world models can produce visually plausible but physically wrong rollouts. This failure is structural; distinct physical systems can look identical yet diverge under intervention. We expose this problem with controlled benchmarks that fix the visible scene while varying latent physics. We show that such models may recommend infeasible actions, mispredict interaction outcomes, or certify unsafe behavior. We argue that embodied AI requires world models that identify the simplest physical abstraction sufficient to answer an intervention query. Such a model comprises modular components, including environment representation, latent state and parameter estimation, action specification, interventional dynamics, and query-level response. An autonomous orchestrator should identify the relevant abstraction and compose compatible learned and structured components per query. When closed-form physics is unavailable, uncertain, or costly, the transition model may be analytic, simulated, learned, or hybrid, but it must preserve the structure that determines interventional outcomes. This decomposition makes the model interpretable, its components verifiable, and its outputs auditable against the query. It also provides a design principle for new world models and a feasibility test for existing ones: the right abstraction is not the most detailed model of the world, but the simplest model that preserves the distinctions relevant to the query. We demonstrate this approach on queries that existing systems fail to answer correctly, and outline how an orchestrator can dynamically assemble and adapt physically viable models for planning, control, and verification.

Problem

• world models
• embodied AI
• physical viability
• intervention queries
• latent physics

Innovation

• physically viable world models
• intervention queries
• modular decomposition
• embodied AI
• query-conditioned abstraction