LAMP: Large-Scale Autonomous Mapping and Positioning for Exploration of Perceptually-Degraded Subterranean Environments

The evolution of underground SLAM has transitioned from early 2D scan matching algorithms like Hector SLAM to sophisticated 3D lidar-based systems such as LOAM and Cartographer. These methods have achieved success in structured environments but face limitations in perceptually degraded settings characterized by feature scarcity, repetitive structures, and environmental noise. Multi-robot SLAM has been explored to improve coverage and efficiency, with approaches like graph-based fusion and loop closure detection. However, challenges in robustness, outlier rejection, and large-scale map consistency persist. Recent advances incorporate robust optimization and sensor fusion, yet the unique conditions of subterranean environments demand further innovations to handle drift, aliasing, and communication constraints effectively.

The core challenge in underground SLAM is maintaining accurate localization and mapping over large, feature-scarce, and perceptually ambiguous environments. Repetitive tunnel structures cause false loop closures, leading to map distortions. Sensor noise, environmental conditions like dust and water vapor, and limited communication exacerbate these issues. Existing solutions often suffer from high false positive rates or drift accumulation, hindering reliable autonomous operation. Addressing these problems requires robust loop closure validation, outlier rejection, and scalable multi-robot data fusion techniques capable of functioning under adverse conditions.

This work introduces several key innovations: 1) a lidar front-end based on scan-to-submap matching with GICP, providing more accurate odometry; 2) a multi-robot pose graph optimization framework that fuses data centrally; 3) an ICM-based outlier rejection algorithm that filters false loop closures, significantly reducing map distortion; 4) support for multi-sensor fusion, including RGB-D artifact detection, enriching the map with semantic data. These innovations collectively enhance the robustness, accuracy, and scalability of underground SLAM, enabling reliable operation in environments with high perceptual aliasing and environmental noise.

• �� The lidar front-end performs scan-to-scan matching using GICP, initialized with odometry, and scan-to-submap matching for improved accuracy. • Loop closure detection compares current scans with nearby historical scans within a 10m radius, using fitness scores to filter false positives. • Outlier rejection employs ICM, which checks the mutual consistency of loop closures by analyzing pose cycle errors and pairwise compatibility, removing inconsistent matches. • Multi-robot data fusion occurs at a central server, where individual pose graphs are merged, and inter-robot loop closures are identified via shared landmarks or fiducials. • Human operators can manually add or correct loop closures through visualization tools like rviz, further improving map quality. • The system is validated through extensive field tests in underground mines, demonstrating high success rates and low map errors.

Field tests involved six underground mines across the US, with environments ranging from coal to gold mines, covering over 1400 meters. Robots equipped with Velodyne Puck lidar and RGB-D cameras explored these sites, with data transmitted to a central base station. Metrics included loop closure success rate, map error, and drift percentage. Ablation studies compared scan-to-scan versus scan-to-submap odometry, showing the latter's superior accuracy. The impact of ICM was quantified by error reduction and false positive elimination. Multi-robot experiments demonstrated effective map fusion, reducing overall errors and improving localization consistency. Results validated the system's robustness in complex, repetitive, and noisy environments.

LAMP achieved over 95% loop closure success in challenging underground environments, with map errors below 0.2 meters. The drift rate was maintained at approximately 1% of total traveled distance, significantly better than traditional methods. Multi-robot fusion reduced trajectory errors by 30%, and ICM decreased false loop closures by 20%. Artifact localization accuracy improved with semantic detection, supporting mission-critical tasks. These results underscore the system’s capability to produce accurate, consistent maps in environments with high perceptual aliasing and environmental noise.

The system is suited for underground rescue missions, resource exploration, tunnel inspection, and extraterrestrial subterranean exploration. It operates without GPS, relying on lidar and multi-sensor fusion, supporting autonomous navigation and mapping in GPS-denied environments. Its scalability and robustness make it applicable to large-scale, complex underground networks, facilitating scientific research, disaster response, and space missions where reliable autonomous mapping is essential.

Performance degrades in environments with water vapor or reflective surfaces, impairing lidar accuracy. Communication constraints in deep tunnels limit data exchange, affecting multi-robot fusion. Persistent structural repetition and occlusion still pose challenges for reliable loop closure detection. Future work will focus on integrating inertial sensors, deep learning-based feature extraction, and decentralized architectures to address these issues and improve environmental robustness.