OOWM: Structuring Embodied Reasoning and Planning via Object-Oriented Programmatic World Modeling

A research team led by Hongyu Chen has introduced the Object-Oriented World Modeling (OOWM) framework, a novel approach to solving the inherent limitations of text-based reasoning for embodied AI tasks like robotics. Standard Chain-of-Thought prompting in LLMs relies on linear natural language, which struggles to explicitly represent the complex state-space, object hierarchies, and causal dependencies needed for robust planning in the physical world. OOWM addresses this by fundamentally redefining the world model as an explicit symbolic structure, W = ⟨S, T⟩, consisting of a State Abstraction (S) and a Control Policy (T) that defines state transitions.

To materialize this model, OOWM borrows from software engineering, using Unified Modeling Language (UML) diagrams. It employs Class Diagrams to ground visual perception into rigorous object hierarchies and Activity Diagrams to operationalize high-level plans into executable control flows. The team also developed a three-stage training pipeline that combines Supervised Fine-Tuning with a novel Group Relative Policy Optimization (GRPO) method. This pipeline uses sparse, outcome-based rewards from a final executed plan to implicitly optimize the underlying object-oriented reasoning structure, enabling effective learning without dense step-by-step annotations.

Extensive evaluations on the MRoom-30k benchmark demonstrate OOWM's significant advantages. The framework substantially outperforms unstructured textual reasoning baselines across key metrics, including planning coherence, execution success rate, and the structural fidelity of the generated world models. This establishes OOWM as a compelling new paradigm, moving AI reasoning from flexible but ambiguous text toward structured, programmatic representations that are more suitable for reliable interaction with the physical world.