A Distributed Multi-UGV Exploration Framework With Loop-Aware Planning and Descriptor-Aided Localization in Resource-Limited Environments

The evolution of autonomous exploration has transitioned from single-robot SLAM systems like ORB-SLAM and RTAB-Map to multi-robot frameworks aiming for scalability and robustness. Early multi-UGV systems relied on centralized architectures, which faced bottlenecks in communication and computation, especially in GPS-denied environments. Distributed SLAM approaches, such as iSAM2 and decentralized pose graph optimization, addressed some scalability issues but struggled with reliable loop closure detection under viewpoint variations. Traditional descriptors like Scan Context and LiDAR-Iris provided some invariance but lacked robustness in large yaw shifts or dynamic scenes. Recent deep learning-based descriptors, such as OverlapTransformer, improved recognition but at higher computational costs. Path planning strategies like frontier-based exploration and hierarchical task allocation have been developed to improve coverage efficiency, yet they often treat loop closures passively, missing opportunities for active refinement. Overall, the field has made significant progress but still faces challenges in achieving robust, real-time, resource-efficient multi-UGV exploration in complex, unknown environments.

The core challenge in multi-UGV exploration within resource-limited, GPS-denied environments is maintaining accurate, globally consistent localization without relying on centralized infrastructure. This involves overcoming pose drift caused by environmental ambiguities, viewpoint changes, and limited communication bandwidth. Traditional loop closure detection methods often fail under severe yaw and lateral shifts, leading to map misalignments and redundant coverage. Additionally, existing path planning approaches lack active utilization of loop information, resulting in inefficiencies and increased exploration time. The problem becomes more complex when considering the need for decentralized operation, real-time performance, and minimal communication overhead, all while ensuring environmental coverage and localization robustness. Addressing these issues requires innovative descriptors, active loop screening, and integrated hierarchical planning strategies.

This work introduces several key innovations:

1) Spectral-guided range image prealignment: By leveraging spectral features and gradient saliency, the system achieves viewpoint-invariant descriptors that are robust to large yaw and lateral shifts, enabling reliable cross-UGV place recognition.

2) Uncertainty-aware loop selection: The framework scores candidate loops based on pose uncertainty and utility metrics derived from covariance propagation, actively selecting high-value loops as planning anchors.

3) Loop-informed hierarchical planning: Integrates loop information into global task allocation via MDVRP and local path refinement through TSP, ensuring efficient coverage and robust localization.

4) Asynchronous distributed optimization: Employs decentralized pose graph updates with minimal communication, maintaining global consistency in resource-constrained environments.

These innovations collectively enable a highly robust, efficient, and scalable multi-UGV exploration system suited for complex unknown terrains.

• �� Sensor input: Each UGV employs a rotating 3D LiDAR (e.g., RoboSense Helios-16) to acquire dense point clouds, which are converted into multi-elevation range images.
• �� Descriptor extraction: Spectral cues are computed by summing low-frequency Fourier magnitudes along elevation, combined with median vertical range gradients, to generate saliency maps. The range image is circularly shifted to align with maximum saliency, then passed through a lightweight Swin Transformer-based network to produce a 256-dimensional global descriptor.
• �� Cross-UGV loop detection: Descriptors are stored in local KD-trees; upon new keyframes, nearest neighbors are retrieved, gating matches by similarity. Verified matches undergo Fast-GICP alignment, and confirmed loop closures are asynchronously inserted into a decentralized pose graph optimized via iSAM2.
• �� Map representation: A sparse obstacle-aware topological graph is constructed within a local planning horizon, with nodes representing traversable voxel centers and edges representing collision-free paths. Subgraph fusion occurs asynchronously through descriptor-based node matching.
• �� Hierarchical planning: Frontier voxels are clustered to identify exploration targets. Candidate viewpoints are sampled on a spherical shell, scored by information gain, and assigned via MDVRP. Path refinement employs TSP on the selected viewpoints, incorporating loop closures as waypoints.
• �� Loop selection: Candidate loops are scored based on pose covariance and Euclidean distance, with top-K loops chosen as planning anchors, actively influencing path planning and localization.

• �� Platforms: Experiments utilize a fleet of UGVs equipped with RoboSense Helios-16 LiDAR, IMU, onboard GPU (Nvidia RTX 4080), and onboard computation, both in real-world environments and Gazebo simulation.
• �� Datasets: Evaluation uses KITTI and Mulran datasets for loop detection robustness, covering diverse scenarios such as dynamic occlusion, large yaw variation, and reverse revisits.
• �� Metrics: Performance is measured via AR@1 and AR@1%, absolute trajectory error (ATE), communication volume, exploration time, and path length.
• �� Baselines: Comparisons include Fast-LIO2, DCL-SLAM (LiDAR-Iris), Kimera-Multi, and other recent descriptors.
• �� Hyperparameters: Descriptor dimension set to 256, top-K loop candidates for planning, and path optimization via A* and TSP algorithms.
• �� Validation: Both simulation and real-world tests confirm the system’s robustness, efficiency, and scalability across environments with different structural complexities.

• �� Loop closure detection: Achieves AR@1 of 89.9% and AR@1% of 95.5%, outperforming LiDAR-Iris (AR@1 84.57%) and deep learning descriptors like OverlapTransformer (AR@1 80.67%) especially under large viewpoint shifts.
• �� Localization accuracy: Average trajectory error drops below 0.45 meters across complex environments, outperforming standalone Fast-LIO2 and other SLAM systems.
• �� Communication efficiency: Total data exchange volume is reduced by approximately 40%, enabling operation under bandwidth constraints.
• �� Exploration efficiency: Overall exploration time is shortened by 15%, and total travel distance by 14%, demonstrating the effectiveness of active loop utilization and hierarchical path planning.
• �� Scalability: The framework maintains high performance with increasing number of UGVs, validating its suitability for large-scale deployments.

• �� Immediate applications include disaster response in GPS-denied urban or underground environments, underground mining, tunnel inspection, and planetary surface exploration.
• �� The system requires high-quality LiDAR sensors, stable wireless communication, and onboard computational resources.
• �� Industry impact involves improved autonomous navigation robustness, reduced operational costs, and enhanced scalability for multi-robot deployments in complex terrains.

• �� Dependence on high-performance sensors: In environments with poor sensor data or dynamic clutter, descriptor robustness may decline.
• �� Computational load: Descriptor extraction, loop verification, and distributed optimization pose challenges for real-time operation in extremely large or highly dynamic environments.
• �� Environmental complexity: Highly dynamic scenes with frequent occlusions or moving obstacles may impair loop detection and map consistency, requiring further robustness enhancements.