AGWM: Affordance-Grounded World Models for Environments with Compositional Prerequisites

Researchers propose AGWM (Affordance-Grounded World Models), a machine learning framework that improves how AI agents understand which actions are executable in dynamic environments by explicitly tracking prerequisite dependencies. The approach addresses a fundamental limitation in conventional world models that fail to account for how actions reshape the availability of future actions, reducing multi-step prediction errors and improving generalization.

Standard world models in reinforcement learning operate under a critical assumption: that the relationship between states and actions remains stationary. AGWM challenges this assumption by recognizing that interactive environments are fundamentally compositional—actions become available or unavailable based on whether their prerequisites are satisfied. This distinction matters because traditional models internalize spurious correlations between actions and outcomes while ignoring the logical preconditions that actually govern executability.

The paper addresses a real problem in sequential decision-making. When an agent imagines rolling out trajectories multiple steps into the future, each prediction compounds errors from the previous step. If the model incorrectly assumes an action is executable when its prerequisites aren't met, subsequent predictions cascade from an impossible state. By representing affordance structure as a directed acyclic graph (DAG) of dependencies, AGWM enables the model to dynamically update which actions remain viable.

The implications extend beyond academic curiosity. More accurate world models improve planning in robotics, game-playing agents, and any domain where action feasibility changes over time. Better multi-step predictions reduce sample complexity during training, which translates to faster learning. The framework's improved generalization to novel configurations suggests it captures genuine causal structure rather than memorizing training correlations.

Future work should examine scalability to complex real-world environments with thousands of interdependent actions. Integration with large language models for abstract prerequisite reasoning could accelerate adoption in practical applications. The interpretability benefits of explicit DAG representation may also enable easier debugging and human-AI collaboration in planning systems.

• →AGWM explicitly models action prerequisites as a DAG to track dynamic executability in interactive environments
• →Conventional world models fail at multi-step prediction because they ignore preconditions and assume stationary transition functions
• →The approach reduces prediction error compound effects and improves generalization to novel configurations
• →Explicit affordance structure representation enhances interpretability by revealing causal relationships between actions
• →Practical applications include robotics, game AI, and planning systems where action feasibility changes over time