⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Dyn-O: Building Structured World Models with Object-Centric Representations

arXiv 25.7 2025 59.3 method

TLDR

Dyn-O builds object-centric world models from pixels, improving dynamics prediction and compositional generalization on Procgen games.

Reasoning

The paper introduces a novel object-centric world model that outperforms DreamerV3 in rollout prediction on Procgen games, with strengths in decoupling dynamics-agnostic and dynamics-aware features for finer-grained manipulation. However, it is only evaluated on simulated game environments, limiting claims of real-world applicability, and the visual complexity of Procgen may still be less than real-world scenes.

Read-first score

Read-first score 59.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 58.

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%
82.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 25%
40

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 657.

Keyword Scores

world model
10
world dynamics prediction
10
generative world model
9
world simulator
8
model-based reinforcement learning world model
8
video world model
7
interactive world model
6

Deep Analysis

Innovations

  • Enhanced structured world model built upon object-centric representations that works in complex visual settings (Procgen games with diverse textures and cluttered scenes)
  • Decoupling object-centric features into dynamics-agnostic and dynamics-aware components for finer-grained manipulation and generation of more diverse imagined trajectories

Methodology

Dyn-O is an object-centric world model that learns representations from pixel observations. It decouples object-centric features into dynamics-agnostic and dynamics-aware components to improve dynamics modeling and trajectory diversity. The model is evaluated on Procgen games against DreamerV3 using rollout prediction accuracy.

Key Results

On Procgen games, Dyn-O outperforms DreamerV3 in rollout prediction accuracy. Additionally, the decoupling enables finer-grained manipulation and generation of more diverse imagined trajectories.

Resources

Paper

Tags