TACS-Graphs: Traversability-Aware Consistent Scene Graphs for Ground Robot Localization and Mapping

Jeewon Kim, Minho Oh, Hyun Myung

Introduction

3D scene graphs have emerged as a powerful hierarchical representation for indoor environments, mapping not only geometric structures but also the semantic relationships between objects and rooms. These graphs are vital for high-level robot task planning and navigation. However, existing state-of-the-art methods often struggle with consistency. Relying on voxel-based representations like Octomaps or Euclidean Signed Distance Fields (ESDFs) frequently leads to segmentation errors that degrade the robot's localization performance.

This paper introduces TACS-Graphs (Traversability-Aware Consistent Scene Graphs), a framework designed to improve the reliability of 3D scene graphs by grounding room segmentation in the robot's physical traversability—defining space by where a robot can actually move rather than mere geometric proximity.

The Problem of Inconsistent Scene Graphs

Traditional 3D scene graph generation faces two primary consistency hurdles:

  1. Under-segmentation: Geometry-based methods often fail to distinguish between traversable rooms and non-traversable voids, such as the space behind a railing or an open balcony. This leads to inaccurate room boundaries.
  2. Over-segmentation: Sensor noise and high computational demands can cause a single physical room to be fragmented into multiple overlapping nodes. This results in a "messy" graph that lacks temporal stability, making it difficult for the robot to recognize the same location across different visits.

When the underlying graph is inconsistent, loop closure detection—the process of recognizing a previously visited location to correct odometry drift—becomes unreliable, leading to significant errors in the robot's estimated trajectory.

The TACS-Graphs Framework

TACS-Graphs addresses these issues through three integrated modules:

1. Traversability-Aware Room Segmentation

Instead of relying solely on raw voxels, TACS-Graphs utilizes a Boundary-constrained Topological Map Synthesis (B-TMS) technique to extract a "traversable graph." By defining "free space" based on the robot's physical ability to navigate, the system naturally filters out non-traversable areas (like spaces behind half-walls or railings) that typically confuse geometry-only segmenters.

2. Room Merging (RM) Module

To resolve over-segmentation, the RM module evaluates newly detected room components against the existing graph. If two components share overlapping traversable space and their centroids align along orthogonal axes (x or y), they are merged into a single, unified room node. This ensures the graph remains stable even in the presence of sensor noise.

3. CoSG-LCD: Consistent Scene Graph-leveraging Loop Closure Detection

By maintaining a stable room layer, TACS-Graphs can optimize loop closure detection. The CoSG-LCD module employs a coarse-to-fine strategy: it first identifies the specific room the robot is in and then searches for matching keyframes within that room. This hierarchical approach is faster and more resilient to large initial pose errors than global search methods.

Experimental Evaluation

The framework was evaluated using public benchmarks (TIERS LiDAR and S-Graphs datasets) and real-world data from challenging environments like the KAIST-E3 building and KEPCO power stations.

Key performance metrics include:

  • Segmentation Consistency: TACS-Graphs achieved a Dice Coefficient Score (DCS) of 94.2% on the Indoor07 dataset. In contrast, the baseline S-Graphs+ failed to segment the environment correctly, resulting in a 0% score.
  • Localization Accuracy: TACS-Graphs demonstrated superior precision with a weighted Absolute Trajectory Error (wATE) of 1.37m, significantly lower than S-Graphs+ (2.38m) and Fast-GICP (3.02m).
  • Computational Efficiency: The reduction in redundant room nodes led to lower factor graph optimization times ($t_{FGO}$), streamlining the mapping process.

Conclusion and Future Work

TACS-Graphs demonstrates that using physical constraints—specifically traversability—as a semantic prior creates more consistent and useful 3D scene graphs. This consistency creates a virtuous cycle: better mapping enables more accurate loop closures, which in turn produces more accurate poses for further mapping.

Currently, the system relies on a Manhattan World Assumption, assuming walls are rectilinear. Future work will focus on generalizing the segmentation logic to handle non-orthogonal layouts and unstructured environments, as well as extending the framework to transition seamlessly between indoor and outdoor spaces.

Summary updated

BibTeX
@inproceedings{kim2025tacs,
  author = {Jeewon Kim and Minho Oh and Hyun Myung},
  title = {{TACS-Graphs: Traversability-Aware Consistent Scene Graphs for Ground Robot Localization and Mapping}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11247779},
}
Figure