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

Aether: Geometric-Aware Unified World Modeling

arXiv 25.3 2025 63.4 method, system

TLDR

Aether unifies geometric reconstruction and generative modeling for world models, enabling zero-shot synthetic-to-real generalization in 4D reconstruction, video prediction, and visual planning.

Reasoning

The paper presents a novel unified framework that integrates multiple capabilities, with strong claims of zero-shot generalization. Strengths include a clear problem formulation and synergistic learning approach; weaknesses include lack of explicit real-world training data and potential overclaiming of generalization without detailed empirical validation in the abstract.

Read-first score

Read-first score 63.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.

Recency 8%
86.7

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

Topical relevance 42%
70

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%
70

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,metric

Reproducibility 25%
38

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=github

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 181.

Keyword Scores

world model
9
generative world model
8
video world model
8
world dynamics prediction
8
interactive world model
7
model-based reinforcement learning world model
5
world simulator
4

Deep Analysis

Innovations

  • Unified framework integrating 4D dynamic reconstruction, action-conditioned video prediction, and goal-conditioned visual planning
  • Task-interleaved feature learning for synergistic knowledge sharing across reconstruction, prediction, and planning objectives
  • Zero-shot synthetic-to-real generalization without real-world training data
  • Camera trajectories as geometry-informed action spaces for action-conditioned prediction and visual planning

Methodology

Aether jointly optimizes three core capabilities—4D dynamic reconstruction, action-conditioned video prediction, and goal-conditioned visual planning—through task-interleaved feature learning. It builds upon video generation models and uses camera trajectories as geometry-informed action spaces, training exclusively on synthetic data.

Key Results

The framework achieves zero-shot synthetic-to-real generalization in action following and reconstruction tasks, with reconstruction performance comparable to or better than domain-specific models despite never observing real-world data during training.

Limitations

  • Reliance solely on synthetic training data may limit generalization to complex real-world scenarios not captured in simulation
  • Zero-shot generalization claims may not hold for all tasks or environments beyond those tested
  • Use of camera trajectories as action spaces may restrict applicability to domains where such trajectories are not available or meaningful
  • No explicit evaluation metrics or comparisons for video prediction and visual planning tasks are provided in the abstract

Resources

PaperProject

Tags