Latent-Space Autoregressive World Model for Efficient and Robust Image-Goal Navigation

arXiv 25.11 2025 59.2 method, application

TLDR

A latent-space autoregressive world model using DINOv3 for efficient, robust image-goal navigation, validated on benchmarks and a real humanoid robot.

Reasoning

The paper presents a novel approach that avoids pixel-level reconstruction by operating in a semantic latent space, achieving state-of-the-art results and real-world deployment. However, the abstract lacks detailed discussion of limitations or comparisons to other latent-space methods.

Read-first score

Read-first score 59.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

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 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 205.

Keyword Scores

world model

world dynamics prediction

interactive world model

model-based reinforcement learning world model

generative world model

world simulator

video world model

Deep Analysis

Innovations

Operating entirely in latent space for world model prediction and planning, bypassing pixel-level reconstruction
Using DINOv3 as a pre-trained representation encoder with specialized mechanisms to integrate action signals and historical context
Autoregressive world model that predicts future latent states conditioned on actions and history
Real-world deployment on a Unitree G1 humanoid robot demonstrating efficiency and robustness

Methodology

The proposed ReL-NWM leverages a pre-trained DINOv3 encoder to produce high-level semantic latent representations. An autoregressive world model then predicts future latent states conditioned on action signals and historical context, all within the latent domain, avoiding explicit pixel-level reconstruction. The model is trained and evaluated on multiple benchmarks for trajectory prediction and image-goal navigation, with comparisons to baselines.

Key Results

The method achieves state-of-the-art trajectory prediction and image-goal navigation performance on multiple benchmarks. Real-world deployment on a Unitree G1 humanoid robot confirms its efficiency and robustness in practical navigation scenarios.

Resources

Paper