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

2026-05-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

The Affordance-Grounded World Model (AGWM) addresses limitations in standard world models that fail to track dynamic action executability in environments with compositional prerequisites, termed Structure-Changing (SC) events. Conventional models often internalize action-outcome correlations without considering preconditions, leading to compounding prediction errors in multi-step rollouts and poor generalization to novel configurations. AGWM learns an abstract affordance structure, represented as a Directed Acyclic Graph (DAG) of prerequisite dependencies, to explicitly track the dynamic executability of actions. It incorporates an SC Classifier to detect structure changes and a Dynamic Affordance Graph that enforces a frontier-mask constraint, ensuring only feasible actions are considered during imagination. Experiments in game-based simulated environments demonstrate AGWM's effectiveness, achieving lower multi-step prediction error, better generalization to novel configurations, and improved interpretability compared to baseline models.

Key takeaway

For research scientists developing model-based reinforcement learning agents in compositional domains, AGWM offers a robust solution to the challenge of dynamic action executability. By explicitly modeling affordance structures and their evolution, you can significantly reduce multi-step prediction errors and enhance generalization to unseen rule combinations. Consider integrating AGWM's self-evolving graph and frontier-mask constraint to improve the safety and reliability of your autonomous agents by ensuring imagined trajectories adhere to feasible action sequences.

Key insights

AGWM explicitly tracks dynamic action executability via a self-evolving affordance graph, reducing prediction errors in complex environments.

Principles

• Explicitly model action preconditions.
• Affordance graphs improve generalization.
• Self-evolving graphs adapt to novel rules.

Method

AGWM augments a recurrent world model with an SC Classifier and a Graph Predictor, using a Graph Encoder to embed affordance structure into the GRU input and decoder, and enforcing a frontier-mask constraint.

In practice

• Use AGWM for compositional RL tasks.
• Apply frontier-mask for tech-tree environments.
• Condition decoder on graph embedding.

Topics

• Affordance-Grounded World Models
• Structure-Changing Events
• Dynamic Affordance Graph
• Model-Based Reinforcement Learning
• Compositional Generalization

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• WALL-WM: Carving World Action Modeling at the Event Joints — Shifting VLA learning from fixed-length chunks to semantically coherent action events resolves granularity mismatch and improves WAM generalization.
• SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale — Dynamic, graph-based structural retrieval significantly enhances LLM agent skill selection and performance over static methods.
• The Sequence Knowledge #800: Not All World Models are Created Equal — Model-Free RL agents learn reactive behaviors without internal environmental understanding.
• ADAPT: Benchmarking Commonsense Planning under Unspecified Affordance Constraints — Embodied agents require dynamic affordance reasoning to adapt to unspecified, changing real-world object constraints.
• BRICKS-WM: Building Reusability via Interface Composition Kinetics for Structured World Models — BRICKS-WM enables modular, reusable world models in MBRL by functionally separating agent and environment dynamics.
• GraphWorld: Long-Horizon Planning with World Models for End-to-End Autonomous Driving — GraphWorld leverages latent world models and interaction graphs to achieve robust, long-horizon planning for end-to-end autonomous driving.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.