Multi-Type Map Construction via Semantics-Aware Autonomous Exploration in Unknown Indoor Environments

Jianfang Mao, Yuheng Xie, Si Chen, Zhixiong Nan, Xiao Wang

Abstract

This paper proposes a novel semantics-aware autonomous exploration model to handle a long-standing issue in mobile robotics: mainstream RRT (Rapid-exploration Random Tree) based models often cause robots to switch frequently between different regions, leading to excessively repeated exploration. Our proposed semantics-aware model encourages a mobile robot to fully explore its current region before moving to the next, avoiding redundancy and accelerating the mapping process. The core innovation lies in optimizing the sampling point selection mechanism and frontier point evaluation function using regional semantic information. Furthermore, unlike existing methods that construct only one or two map types, our model simultaneously generates four: point cloud, occupancy grid, topological, and semantic maps. Experimental results in simulated environments show that our model achieves a 33.0% reduction in exploration time and a 39.3% reduction in trajectory length while maintaining a >98% exploration rate.

Introduction

Autonomous exploration is the process by which a robot enters an unknown space and builds a map while deciding where to go next. However, current state-of-the-art methods suffer from two major flaws. First is the "Ping-Ponging" effect: algorithms often select the next goal based on the largest frontier of unknown space, causing a robot to abandon a half-explored room for a nearby hallway, only to backtrack later. Second is Map Monoculture: most systems generate only a single map type (e.g., a 2D grid), whereas robots actually need 3D data for localization, topological data for path planning, and semantic data for high-level task understanding.

Semantic Region Aware Navigation

To prevent premature region switching, the proposed method introduces two key modifications to standard RRT exploration:

  • Dynamic Sampling: The system uses a probability mechanism to select sampling points. Early in a room's exploration, the robot is heavily biased toward sampling points within its current semantic region. As time passes, this bias relaxes, allowing the robot to move to the next area once the current one is likely finished.
  • Region-Based Rewards: When evaluating potential "frontier points" (the boundary between known and unknown space), the algorithm adds a positive reward if the point is in the robot's current room and a penalty if it is outside.
Comparison of standard RRT sampling vs. semantic region aware sampling logic.

Multi-Type Map Construction

The system builds four maps simultaneously using a shared coordinate system to ensure consistency and computational efficiency:

  1. 2D Occupancy Grid: Built via Cartographer SLAM for basic navigation.
  2. 3D Point Cloud: Built via LIO-SAM using Lidar and IMU data for detailed environmental modeling.
  3. Topological Map: Created by extracting a skeleton from the 2D map to identify paths and nodes.
  4. Semantic Map: Uses image processing techniques to segment the environment into distinct rooms and assign unique IDs.
The four types of maps generated simultaneously: (1) 2D occupancy, (2) topological, (3) 3D point cloud/trajectory, and (4) semantic.

Key Results

The proposed method significantly outperformed classical RRT-based exploration in Gazebo simulations across various house and office layouts:

  • Efficiency Gains: The model achieved a 50.7% reduction in total exploration time in medium-sized environments and a 33.0% reduction overall.
  • Trajectory Shortening: Total distance traveled was reduced by up to 48.1%.
  • High Coverage: The robot maintained an exploration rate of >98%, proving that being "region-aware" does not lead to missing spots.
  • Real-time Performance: Map updates for 2D and topological maps stayed around 1 second, while 3D point clouds updated in just 0.2 seconds.

Conclusion

By enforcing a "one room at a time" policy, the robot avoids the long, repetitive paths that plague standard frontier-based exploration. While highly effective indoors, the current method relies on structural features like walls and doors for segmentation. Future work will focus on extending this logic to outdoor scenarios, where boundaries are more fluid, and handling dynamic obstacles during the mapping process.

Summary updated

Abstract

This paper proposes a novel semantics-aware autonomous exploration model to handle the long-standing issue: the mainstream RRT (Rapid-exploration Random Tree) based exploration models usually make the mobile robot switch frequently between different regions, leading to the excessively-repeated explorations for the same region. Our proposed semantics-aware model encourages a mobile robot to fully explore the current region before moving to the next region, which is able to avoid excessively-repeated explorations and make the exploration faster. The core idea of semantics-aware autonomous exploration model is optimizing the sampling point selection mechanism and frontier point evaluation function by considering the semantic information of regions. In addition, compared with existing autonomous exploration methods that usually construct the single-type or 2-3 types of maps, our model allows to construct four kinds of maps including point cloud map, occupancy grid map, topological map, and semantic map. To test the performance of our model, we conducted experiments in three simulated environments. The experiment results demonstrate that compared to Improved RRT, our model achieved 33.0% exploration time reduction and 39.3% exploration trajectory length reduction when maintaining >98% exploration rate.
BibTeX
@misc{mao2024multi,
  author = {Jianfang Mao and Yuheng Xie and Si Chen and Zhixiong Nan and Xiao Wang},
  title = {{Multi-Type Map Construction via Semantics-Aware Autonomous Exploration in Unknown Indoor Environments}},
  year = {2024},
  eprint = {2404.04879},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.04879},
}
Figure