Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator

arXiv 2026 35.3 method, system

TLDR

Image2Sim builds interactive 3D environments from posed RGB-D images using neural Gaussian models, generating 20K scenes and 10M navigation samples to improve embodied navigation.

Reasoning

Strengths include a novel decoupling of spatial anchoring and observation synthesis, enabling scalable data generation and strong benchmark improvements. Weaknesses are reliance on posed RGB-D input and limited discussion of sim-to-real transfer validation.

Read-first score

Read-first score 35.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

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

Topical relevance 25%

52.9

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

Reproducibility 15%

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 341.

Keyword Scores

world simulator

generative world model

world model

interactive world model

world dynamics prediction

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Decoupling 3D spatial anchoring from photorealistic observation synthesis for neural simulation
Feed-forward feature Gaussian model for single-pass scene construction from posed RGB-D images
Geometry-Aware One-Step Pixel Flow model to render high-quality panoramic RGB-D from sparse Gaussian projections
Fully automated embodied data engine that generates high-fidelity observations, actions, and diverse instructions at scale
Scaling neural simulation to 20K interactive scenes and 10M navigation samples, enabling zero-shot real-world transfer

Methodology

Image2Sim constructs 3D environments from posed RGB-D sequences using a feed-forward feature Gaussian model, then renders panoramic RGB-D observations via a Geometry-Aware One-Step Pixel Flow model. It automates generation of navigation data including observations, actions, and instructions, training models entirely in neural environments.

Key Results

Navigation models trained in Image2Sim's neural environments show strong improvements on major benchmarks and transfer effectively to real-world zero-shot settings, using 20K scenes and over 10M synthesized training samples.

Resources

Paper