← Paper Digest

I See You: Initial Multi-Robot LiDAR Registration with Mutual-Observation Checks

Jonathan Lichtenfeld, Oskar von Stryk

November 2026

TL;DR A pre-mission pipeline aligns a team of ground robots into one frame from a single LiDAR snapshot each, and rejects impossible pairwise alignments by checking that the robots do not overlap and that each robot appears where the other's LiDAR should see it. On a real four-robot dataset the checks reject 37.3% of wrong pairwise registrations and only 0.6% of correct ones.

What they do

Search-and-rescue teams of ground robots need a shared reference frame before the mission starts, without markers, GNSS or prior maps. This framework registers robots that start together in a staging area, using one gravity-aligned LiDAR snapshot per robot. Its focus is catching alignments that a global point-cloud matcher returns with confidence but that are physically impossible, which happens in repetitive or sparse scenes.

How it works

  • Pairwise registration. Every robot pair is aligned with KISS-Matcher, searching yaw only since scans are gravity-aligned, then refined with GICP.
  • Plausibility checks. A pose is rejected if the two chassis bounding boxes overlap, or if the predicted box of the other robot is empty while the LiDAR has returns from the space behind it (mutual observation, run in both directions). The checks can only reject a pose, never confirm it.
  • Team assembly. Accepted pairs form a maximum-weight spanning tree weighted by inlier count. Rejected pairs are retried from poses composed along the tree, and all verified edges go into a GTSAM pose graph.
  • Dataset. Four wheeled and tracked UGVs with Ouster and Livox LiDARs; 69 static snapshots across seven indoor and outdoor scenes, with terrestrial-laser-scanner ground truth in four of them.

Results

  • Of the 536 wrong pairwise registrations the matcher accepted, the checks reject 200 (37.3%), and they reject 7 of 1145 correct ones (0.6%). Inlier-count and RMSE cutoffs set to reject the same 200 wrong pairs reject 130 and 11 correct ones.
  • Full team, one run per snapshot: 51 of 69 succeed (73.9%), 14 place a robot wrongly and 4 leave one unregistered. Successful TLS-referenced runs average 5.40 cm and 0.51° error. A run takes 5.60 s on average.
  • Over 598 runs on every robot subset, the checks cut wrong-pose outcomes from 127 to 79. A robust GNC-TLS back end without the checks does not reduce them.

Limitations

The mutual-observation check mainly fires in open, planar scenes (true-positive rate 0.70 on Walkway, 0.03 on Meadow) and abstains in clutter and vegetation. It needs the robots to see each other, known chassis dimensions and calibrated LiDAR extrinsics, and it cannot reject all small errors. The all-pairs registration step limits how far it scales to larger teams.

Figures

From Figure 1: A team of four heterogeneous UGVs at a common staging area (top) and their LiDAR scans aligned into a ...
From Figure 1: A team of four heterogeneous UGVs at a common staging area (top) and their LiDAR scans aligned into a common reference frame by the method (bottom), colored per robot.
From Figure 2: The registration pipeline. Pairwise registrations are filtered by physical plausibility checks, and ac...
From Figure 2: The registration pipeline. Pairwise registrations are filtered by physical plausibility checks, and accepted edges become candidates for a maximum-weight spanning tree weighted by inlier count. Rejected edges are retried with an initial transform composed along the tree; if the retry passes the same checks, it is added to the final pose graph optimization.
From Figure 4: The captured scenes. 1: Lab, 2: Parking, 3: Walkway, 4: Campus, 5: Forest, 6: Meadow, 7: Park, 8: capt...
From Figure 4: The captured scenes. 1: Lab, 2: Parking, 3: Walkway, 4: Campus, 5: Forest, 6: Meadow, 7: Park, 8: capturing the terrestrial laser scanner reference scans.
Abstract

Compared to single-robot deployments, multi-robot teams can significantly improve search-and-rescue (SAR) mission success by covering large areas in parallel. To collaborate effectively, these robots must be registered in a common reference frame. Standard registration methods typically rely on external infrastructure or significant map overlap, but real deployments require infrastructure-free, stationary initial alignment for immediate operation. While global LiDAR point cloud registration addresses this, existing single-scan methods often fail silently in aliased or degenerate environments. Approaches scaling beyond two agents also remain rarely evaluated in real-world scenarios. We propose a centralized, deployment-ready framework for pre-mission multi-robot registration. Our system performs pairwise coarse-to-fine alignment and actively rejects impossible pose hypotheses via deterministic plausibility checks. We validate the framework on a real-world custom dataset of four synchronized robots in various environments with ground truth. The checks reject more than a third of the incorrect pairwise registrations while falsely rejecting less than 1% of the correct ones, and reduce the number of team registrations with a wrong robot pose by 38%. Dataset and code are publicly available.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.