Multi-robot LiDAR SLAM: a practical case study in underground tunnel environments

Multi-robot SLAM aims at localizing and building a map with multiple robots, interacting with each other. In the work described in this article, we analyze the pipeline of a decentralized LiDAR SLAM system to study the current limitations of the state of the art, and we discover a significant source of failures, i.e., that the loop detection is the source of too many false positives. We therefore develop and propose a new heuristic to overcome these limitations. The environment taken as reference in this work is the highly challenging case of underground tunnels. We also highlight potential new research areas still under-explored.

💡 Research Summary

The paper presents a comprehensive study of decentralized multi‑robot LiDAR SLAM in the challenging context of underground tunnels. The authors first identify the dominant source of failure in current multi‑robot SLAM pipelines: an excessive number of false‑positive loop detections caused by perceptual aliasing in repetitive tunnel geometry. To investigate this problem, they create a new high‑fidelity dataset derived from a real tunnel (the Zentrum am Berg underground facility). Four four‑wheeled robots equipped with Ouster OS2‑128 LiDARs traverse four distinct trajectories, generating 261 point clouds with sub‑centimeter accuracy. Unlike existing SubT datasets that assume a single entrance, this dataset includes multiple entry points, making it more realistic for decentralized operation.

Two LiDAR odometry approaches are evaluated: KISS‑ICP, which relies solely on LiDAR scans, and Kinematic‑ICP, which fuses wheel odometry to handle degenerate geometries. Experiments show that KISS‑ICP performs well in richly featured scenes but suffers large absolute trajectory errors (ATE) in the tunnel’s feature‑poor sections, whereas Kinematic‑ICP remains robust thanks to the additional motion constraints.

The multi‑robot SLAM pipeline builds on Swarm‑SLAM concepts. Each robot generates a keyframe every 0.5 m of travel, computes a handcrafted Scan Context descriptor for global place recognition, and searches for matches among keyframes of other robots. When a candidate match exceeds a similarity threshold, the relative transformation is estimated using KISS‑Matcher (FPFH‑based feature matching) together with a graph‑based outlier rejection scheme. The resulting inter‑robot edges are added to each robot’s pose graph, which is optimized with GTSAM. To increase robustness against erroneous loops, the authors incorporate two verification layers: (i) Pairwise Consistency Measurement (PCM) to pre‑filter inconsistent inter‑robot loop closures by finding a maximum‑clique of mutually consistent edges, and (ii) Graduated Non‑Convexity (GNC) as a final optimizer that gradually introduces non‑convexity, thereby avoiding local minima.

A novel, lightweight heuristic is introduced to prune uninformative keyframes that contribute little new geometry. For each keyframe, an oriented bounding box is computed via PCA; the second‑largest dimension (typically the tunnel width) is compared against a manually tuned threshold. If the dimension is below the threshold, the keyframe is deemed “tunnel‑only” and excluded from the graph and from subsequent matching. This simple filter dramatically reduces both place‑recognition and point‑cloud‑registration outliers, as demonstrated by extensive experiments.

Six pairwise robot combinations are evaluated under four configurations: (a) all keyframes without PCM, (b) all keyframes with PCM, (c) tunnel‑filtered keyframes without PCM, and (d) tunnel‑filtered keyframes with PCM. Results (Tables 2‑3) reveal that using all keyframes leads to outlier rates above 30 % and successful map merging in only 2 of 6 cases even after PCM. In contrast, the tunnel‑filtered approach cuts outlier rates to below 5 % before PCM and yields successful merging in 5 of 6 cases after PCM. The combination of PCM and the filtering heuristic thus provides the most reliable performance, while GNC further refines the pose‑graph solution but cannot compensate for a flood of bad loops on its own.

The authors conclude that robust optimization alone is insufficient for practical multi‑robot SLAM in feature‑poor environments; data‑level pruning is essential. Their contributions are threefold: (1) a publicly released, high‑quality underground tunnel dataset with multiple entry points, (2) a simple yet effective keyframe filtering heuristic based on oriented bounding‑box dimensions, and (3) an integrated verification pipeline (PCM + GNC) that substantially improves loop‑closure reliability in decentralized settings. Limitations include the heuristic’s dependence on environment‑specific thresholds and the risk of discarding useful loops if the filter is too aggressive. Future work is suggested in adaptive threshold learning, multimodal sensor fusion (e.g., IMU, vision), and validation on real underground deployments.