Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Active Semantic Mapping and Pose Graph Spectral Analysis for Robot Exploration

Rongge Zhang, Haechan Mark Bong, Giovanni Beltrame

TL;DR

This work integrates semantic mutual information with pose-graph spectral analysis to form an active SLAM framework for robot exploration. By weighting trajectory decisions with both semantic information gain and graph-connectivity-based uncertainty, it achieves improvements in metric-semantic SLAM performance without sacrificing exploration efficiency. The approach leverages a Shannon-Rényi entropy formulation and Laplacian spectral properties to avoid costly full-covariance computations, enabling real-time decision-making. Experimental results on photorealistic Habitat/Matterport data demonstrate up to $38\%$ localization error reduction, $21\%$ map error reduction, and notable semantic IoU gains, highlighting the practical impact for robust autonomous exploration and scene understanding.

Abstract

Exploration in unknown and unstructured environments is a pivotal requirement for robotic applications. A robot's exploration behavior can be inherently affected by the performance of its Simultaneous Localization and Mapping (SLAM) subsystem, although SLAM and exploration are generally studied separately. In this paper, we formulate exploration as an active mapping problem and extend it with semantic information. We introduce a novel active metric-semantic SLAM approach, leveraging recent research advances in information theory and spectral graph theory: we combine semantic mutual information and the connectivity metrics of the underlying pose graph of the SLAM subsystem. We use the resulting utility function to evaluate different trajectories to select the most favorable strategy during exploration. Exploration and SLAM metrics are analyzed in experiments. Running our algorithm on the Habitat dataset, we show that, while maintaining efficiency close to the state-of-the-art exploration methods, our approach effectively increases the performance of metric-semantic SLAM with a 21% reduction in average map error and a 9% improvement in average semantic classification accuracy.

Active Semantic Mapping and Pose Graph Spectral Analysis for Robot Exploration

TL;DR

This work integrates semantic mutual information with pose-graph spectral analysis to form an active SLAM framework for robot exploration. By weighting trajectory decisions with both semantic information gain and graph-connectivity-based uncertainty, it achieves improvements in metric-semantic SLAM performance without sacrificing exploration efficiency. The approach leverages a Shannon-Rényi entropy formulation and Laplacian spectral properties to avoid costly full-covariance computations, enabling real-time decision-making. Experimental results on photorealistic Habitat/Matterport data demonstrate up to localization error reduction, map error reduction, and notable semantic IoU gains, highlighting the practical impact for robust autonomous exploration and scene understanding.

Abstract

Exploration in unknown and unstructured environments is a pivotal requirement for robotic applications. A robot's exploration behavior can be inherently affected by the performance of its Simultaneous Localization and Mapping (SLAM) subsystem, although SLAM and exploration are generally studied separately. In this paper, we formulate exploration as an active mapping problem and extend it with semantic information. We introduce a novel active metric-semantic SLAM approach, leveraging recent research advances in information theory and spectral graph theory: we combine semantic mutual information and the connectivity metrics of the underlying pose graph of the SLAM subsystem. We use the resulting utility function to evaluate different trajectories to select the most favorable strategy during exploration. Exploration and SLAM metrics are analyzed in experiments. Running our algorithm on the Habitat dataset, we show that, while maintaining efficiency close to the state-of-the-art exploration methods, our approach effectively increases the performance of metric-semantic SLAM with a 21% reduction in average map error and a 9% improvement in average semantic classification accuracy.
Paper Structure (22 sections, 25 equations, 5 figures, 2 tables)

This paper contains 22 sections, 25 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Semantic Mapping
  4. Exploration
  5. Active SLAM
  6. Spectral Graph Theory
  7. Preliminaries
  8. Occupancy Grid Mapping and Mutual Information
  9. Graph-based SLAM and Optimality Criteria
  10. Optimality Criteria On the Spectra
  11. Active SLAM Based on Semantic Mutual Information and Pose Graph Spectrum
  12. Semantic Mutual Information
  13. Underlying Pose Graph Spectrum for Optimality Criteria
  14. Utility Function
  15. Autonomous Exploration
  16. ...and 7 more sections

Figures (5)

  • Figure 1: Representation of our problem: a robot explores a Habitat environment, building a 3D semantic map using depth sensors and segmentation, and using it for navigation.
  • Figure 2: Schematic diagram of a decision-making process during the active SLAM. The path of frontier A has shorter travel distance while the path of frontier B will provide a more optimistic pose graph with more co-visibility points and loop closure edges for visual SLAM. The final choice of exploration is frontier B. In the figure, red and pink lines represent pose graphs at different stages. Yellow rectangles are candidate frontiers, some of them have been excluded before the decision due to inaccessibility. Green line represents the global path planned by A* algorithm and pink rectangle represents the robot. (a) The robot plans a path toward frontier A and evaluates the corresponding hallucination graph and semantic mutual information. (b) The robot plans a path toward frontier B and evaluates the corresponding hallucination graph and semantic mutual information. (c) The robot executes a path to frontier B. (d) The final generated pose graph after the robot reaches frontier B.
  • Figure 3: SLAM error results for 10 repetitions of the exploration experiment. (a) Localization error in experiment 1. (b) Localization error in experiment 2. (c) Map error in experiment 1. (d) Map error in experiment 2.
  • Figure 4: Visualization of metric-semantic map results obtained during one exploration experiment. (a) Partially explored metric map. (b) Ground truth metric map. (c) The error (Euclidean distance) in CloudCompare software calculation. (d) Ground truth map visualization for the environment. (e) Partially explored semantic map. (f) Ground truth semantic map.
  • Figure 5: Comparison of exploration results. (a) The curves of explored volume over time in experiment 1. (b) The curves of explored volume over time in experiment 2.