Generation of High-Level Concepts in 3D Scene Graphs via Autoregressive Diffusion

Published September 1, 2026 · Category: Robotics

Overview

arXiv:2608.28733v1 Announce Type: new Abstract: Indoor 3D Scene Graphs (3DSGs) represent environments as multi-layer hierarchies that connect observed geometric primitives (e.g., planes) to higher-level metric-semantic concepts (e.g., rooms, floors, buildings), enabling incremental spatial reasoning for robotic perception and SLAM. However, classical high-level concept generation approaches rely on hand-crafted rules for specific concept classes, while learning-based methods require separate models for graph structure and spatial node features (e.g., centroids), which limits scalability to novel classes and more complex hierarchies. We propose a unified autoregressive diffusion-based graph generative model that jointly learns structure and features, constructing complete 3DSGs bottom-up from observed vertical planes across arbitrary hierarchy depths. Our method consistently surpasses all learning-based and random baselines across 3DSG datasets spanning synthetic scenes, real architectural floor plans, and robotic sensor data, with varying layout complexity and hierarchy depth, and surpasses a one-shot model with oracle access to the target graph size on the largest hierarchy and on real single-floor data. Finally, we propose an adaptation of the Fused Gromov--Wasserstein distance for principled graph-level evaluation of generated 3DSGs against ground truth.

Source

Originally published at arxiv.org.

Related Articles

• Event-triggered Control and Online Learning for Networked Systems under Computational Delays
• Calibration and Comparative Analysis of Forward-Looking Sonar and 3D Sonar for Enhanced Underwater Object Recognition
• Feelium: A Touchable Blimp Body for Aerial Telepresence