TL;DR: SKiD-SLAM is a distributed multi-robot LiDAR SLAM system assembled to fit the budgets of an actual planetary mission — 250 kbps of inter-robot bandwidth, embedded-class onboard compute, and no central server. We deployed it at two planetary analog sites in South Korea: three wheeled rovers mapping the Majang lava-tube cave on Jeju Island, and two quadrupeds carrying out a crater-and-rock science scenario on the KICT planetary analog terrain in Yeoncheon. Under the imposed communication budget, SKiD-SLAM was the only evaluated method that completed distributed map merging.
In future planetary exploration missions, teams of robots will need to collaboratively build maps in unknown environments, from subsurface caves and lava tubes to crater-strewn surface terrains. Such collaboration enables faster and more robust mapping than a single robot can achieve by sharing observations and computation across the team. However, this collaboration typically assumes seamless communication and abundant computing resources, whereas planetary environments impose severe communication limits, constrained onboard computation, and the absence of any central infrastructure.
In this field report, we present SKiD-SLAM, a distributed multi-robot LiDAR SLAM system for such communication-constrained planetary environments, and report its deployment across two representative mission concepts, i.e. (i) cave exploration and (ii) scientific surface exploration. To build the system, we survey the communication and computation budgets of actual planetary missions, and select modules that fit these budgets for each core function of distributed multi-robot LiDAR SLAM. These core functions are recognizing revisited places across robots' maps, estimating inter-robot relative poses, and fusing trajectories without a central server.
Before field deployment, we verify the assembled system on public datasets that resemble planetary environments and in a simulator built on NASA JPL's Open Source Rover. We then validate our SKiD-SLAM in both mission concepts through field deployments in South Korea, where (i) a team of wheeled rovers explores the Majang lava-tube cave on Jeju Island, and (ii) quadruped robots carry out a crater-and-rock science scenario on a planetary analog terrain at the Korea Institute of Civil Engineering and Building Technology (KICT) in Yeoncheon. Across both deployments, our SKiD-SLAM reliably performed distributed multi-robot LiDAR SLAM within the 250 kbps inter-robot communication budget, demonstrating its feasibility under challenging field conditions.
The study is organized around two representative planetary mission concepts, distinguished by terrain, platform, and team composition. Both start from the same condition: N robots are deployed in an unknown environment, each in its own local coordinate frame, with no shared global reference, no GPS, no base station, and no central server. A consistent global map must be built on site, from the robots' own onboard observations.
A team cooperatively maps a confined, lightless lava tube — the kind of subsurface environment that may shelter future habitats, but that no planetary robot has yet entered. The entire team shares a single objective: one consistent map of the passages. Site: the Majang lava-tube cave on Jeju Island, roughly 20 m × 40 m with ceilings from 1 m to 5 m, no natural light, and a floor of loose rock. Team: three wheeled rovers.
A team explores open, cratered terrain with the robots divided across complementary roles rather than sharing one objective: one robot traverses the crater field to map its geometry, while another inspects scattered rocks for science observations. Site: the KICT planetary analog terrain in Yeoncheon, roughly 40 m × 60 m of scattered rocks and impact-crater-like depressions up to 1 m deep. Team: two quadruped robots.
| Sequence | # of robots | Duration [s] | Length [m] | # of keyframes |
|---|---|---|---|---|
| Majang lava-tube cave | 3 | 227.4 / 153.3 / 343.1 | 24.4 / 18.8 / 50.3 | 102 / 52 / 156 |
| KICT planetary analog terrain | 2 | 443.9 / 519.8 | 225.2 / 200.0 | 431 / 468 |
Scale of the field deployments, reported per robot.
Most distributed multi-robot LiDAR SLAM systems are developed and validated on urban and indoor benchmarks, where distinctive geometric structure, high-bandwidth links, and supporting infrastructure are available. Those assumptions do not hold on a planetary surface. Rather than proposing another algorithm, we surveyed what flown missions actually provide, and treat the result as the budget that every module of SKiD-SLAM must respect.
Curiosity's direct-to-Earth link runs at 500 bps–32 kbps; orbital relay reaches 256 kbps–2 Mbps but is visible only around eight minutes per sol. Robot-to-robot links are no more generous: Ingenuity talked to Perseverance over a 20/250 kbps Zigbee radio, and the CADRE rovers share a 1 Mbps mesh across the whole team, with drive-motor interference ruling out continuous transmission. We assume at most 250 kbps per robot within a 1 km range — roughly 1/38,000 of a Wi-Fi 6 link.
The radiation-hardened RAD750 that powers Curiosity and Perseverance runs at 133–200 MHz with 256 MB of RAM. Recent missions turn to COTS processors instead — Ingenuity flew a Snapdragon 801 — but solar power, battery capacity, and thermal limits still cap what is sustainable during a traverse. We target compact onboard computers and validate that every module runs to completion without exhausting onboard memory, with real-time performance as a secondary goal.
Among the candidate methods for each module, we therefore select the one that runs within these communication and memory budgets while remaining as fast and accurate as possible, and assemble the selected modules into SKiD-SLAM.
SKiD-SLAM runs entirely onboard each robot, with no central server. Each robot estimates its own keyframe poses, broadcasts a compact descriptor per keyframe through a shared message pool, and — whenever a teammate's descriptor matches one of its own places — registers the two scans, verifies the resulting link twice, and folds it into its own pose graph. Inter-robot and intra-robot loop closing share the same modules.
Instead of optimizing over all N robots on a central server, each robot α optimizes its own poses plus only those teammate poses that form inter-robot loops with it. Four module choices make that objective solvable within the planetary budgets.
Reason Robots must decide whether they have seen the same place by exchanging descriptors over a kilobit-class link, so descriptor size matters as much as accuracy. Cave walls are geometrically self-similar and their intensity varies little, which defeats descriptors built on raw geometric or intensity features.
Method
SOLiD maps each scan to a 320-byte rotation-invariant global descriptor. Candidate pairs are accepted when
the Euclidean descriptor distance falls below τdist, retrieved through a KD-tree.
SOLiD works across fields of view from 60° to 360° and stays effective under occlusion.
Reason Unlike intra-robot loops, inter-robot loops connect robots observing the same place from markedly different viewpoints. ICP-style local registration has too narrow a convergence basin to handle that discrepancy without an initial guess.
Method KISS-Matcher performs global registration over a wide pose space using Faster-PFH features and k-core-based graph-theoretic pruning, which replaces the maximum-clique bottleneck with a linear-time step. Small-GICP then refines that coarse alignment to fine accuracy.
Reason Laser-ray reflections and repetitive structure produce false-positive links that severely distort the map. When robots traverse an area in opposite directions, overlap is limited and a few large residuals dominate a plain MSE score.
Method
A truncated MSE discards correspondences beyond τMSE and scores only the
overlapping region, accepting a link only when it retains enough inliers
(Ninliers ≥ Nmin) and attains a small enough score.
Reason Measurement-level verification alone cannot catch a false link that happens to align well locally, and the rotation component lies in the non-convex SO(3), so direct optimization from a poor initial guess may converge to a local minimum.
Method Pairwise consistency maximization checks loop closures against each other at the pose level and keeps the maximum consistent clique. The surviving constraints enter a two-stage global–local optimization that first solves for the inter-robot frame transformations, then refines all poses.
Before going to the field, we validated each module choice on public sequences that resemble our target
environments: Tunnel for the cave concept and Moon for the cratered surface.
SOLiD matches the most accurate descriptor (LiDAR Iris) on the PR curves below, while sending 720× less data. Bar lengths are drawn on a log scale because the range spans three orders of magnitude; the printed values are the real descriptor sizes.
Tunnel: 1.164 GB (LiDAR Iris) → 1.617 MB (SOLiD)
Tunnel (left) and Moon (right).
Moon. Local methods collapse; KISS-Matcher is agnostic to the initial rotation.Moon sequence
Local methods are accurate once they converge but only within a narrow basin; a global method converges almost always but leaves a coarse residual. Pairing the two gives both.
Single-robot trajectory accuracy is comparable across methods; the advantage of SKiD-SLAM is in aligning one robot's map to another's, which is the central requirement of distributed multi-robot SLAM.
Tunnel ATE [m] ↓Moon ATE [m] ↓Moon ARE [°] ↓Tunnel, so ARE is reported on Moon only
We compared SKiD-SLAM against DiSCo-SLAM, DCL-SLAM, and Multi-Proxy on the data collected at both sites. At the Majang cave, all three baselines fail to merge the robots' maps into a consistent estimate. At the KICT terrain without bandwidth constraints, DiSCo-SLAM produces misaligned maps while the other three succeed — the module choices validated in the lab carry over to real field conditions.
Multi-robot map merging at the Majang lava-tube cave. Only SKiD-SLAM produces a coherent global map of the cave.
We then tightened the available bandwidth with NetEm, which emulates network conditions by controlling bandwidth, latency, and packet loss. SKiD-SLAM keeps merging the maps consistently as the limit drops from 1,000 kbps to 500 kbps and down to 250 kbps — the inter-robot budget assumed from the mission survey.
SKiD-SLAM at the Majang lava-tube cave under bandwidth limits imposed via NetEm.
Under the same 250 kbps cap, DiSCo-SLAM, DCL-SLAM, and Multi-Proxy all fail to merge their maps, while SKiD-SLAM completes merging. Their average transmitted message sizes and scan drop ratios are higher at every tested limit.
To identify which design choices are responsible for this reliability, we disabled one component at a time and measured its effect on the cave and crater sequences. The decisive module turns out to be site-dependent — except PCM, which matters everywhere.
Component ablation at the Majang lava-tube cave. Removing SOLiD causes a clear failure and removing PCM degrades the merge, whereas the other modules have a smaller effect in this environment.
On the open crater terrain, Scan Context still recognizes inter-robot places and the merge succeeds: the terrain provides distinct height variation along each azimuth. The cave removes that cue — narrow, enclosed geometry with repetitive walls and little vertical structure — and the merge collapses.
Without global registration, Small-GICP has no reliable coarse alignment to start from. On the crater terrain the descriptor alone underconstrains the alignment and the two maps never converge; in the cave, removing it has a smaller effect.
Both deployments used homogeneous teams sharing a near-identical sensor viewpoint, so the truncated MSE shows no visible effect at either site. Its benefit appears once the inter-robot viewpoint diverges, as in heterogeneous ground–aerial teams.
Removing PCM degrades the merge in the cave and on the crater terrain alike. The effect is most pronounced on the crater terrain, where the larger map scale admits more candidate loop closures and raises the chance that a false positive slips past geometric verification.
Cropping each robot's scan to a forward-facing FOV of 36°, 72°, or 108° — 10%, 20%, and 30% of a full 360° scan — still yields a consistent global map at every tested FOV, including the narrowest. This suits the narrow-FOV solid-state LiDARs attractive for close-range planetary survey.
SKiD-SLAM merges maps with a 2 ms delay and under 50% packet loss, the latter thanks to redundant keyframe exchange. It fails at a 200 ms delay or 75% packet loss — the system is designed for local robot-to-robot coordination, not Earth-to-Mars operation.
The complete pipeline ran onboard an Intel NUC 11 Pro with no external compute. Average per-keyframe processing stays around 100–200 ms, with peaks occasionally exceeding 500 ms during inter-robot loop closures.
Cross-platform inter-robot loop closure remains unverified in the field, since each site used a homogeneous team. The outlier-rejection thresholds are selected empirically rather than derived, and the memory footprint grows with trajectory length.
Beyond the quantitative results, the field deployments surfaced operational lessons that benchmark evaluation does not expose.
Staying within the bandwidth budget is necessary but not sufficient. In the Majang cave, curved passages and self-similar walls repeatedly placed robots out of line of sight, so the link dropped regardless of the nominal data rate, and the shared wireless medium was subject to contention from teleoperation links and nearby equipment. Budget compliance and link robustness are distinct requirements, and both must be met — which is why ARCHES deployed a dedicated relay rover, and why the CADRE rovers cannot transmit continuously.
Our deployments reproduce the geometry of planetary caves and cratered terrain, but not the physics: full Earth gravity instead of reduced gravity, damp rock instead of fine abrasive electrostatically charged regolith, and a breathable atmosphere instead of thermal cycling in near-vacuum. Validating a perception system for those conditions needs thermal-vacuum and dust-ingress testing that no analog site can provide.
Our pipeline relies solely on LiDAR geometry to stay light on embedded hardware. Because the Majang lava tube receives no natural light, illumination-independent modalities — thermal imaging, ground- penetrating radar — are a natural complement. We recorded thermal imagery alongside the LiDAR data and release it so the community can test whether such cues help.
Multi-minute communication delays and extended blackouts rule out remote intervention, so a real mission demands more autonomy than ground supervision can supply. Coupling geometric distributed SLAM with learned, multimodal priors — and with heterogeneous ground–aerial teams sharing one map — is, in our view, the more compelling long-term direction.
@article{kim2026skid,
title={SKiD-SLAM: Resource-Aware Distributed Multi-Robot LiDAR SLAM for Planetary Missions with Field Deployments at Analog Sites},
author={Kim, Hogyun and Kim, Juwon and Cho, Dongjin and Choi, Jiwon and Yang, Geonmo and Jeong, Seokhwan and Lee, Jiyun and Park, Miryeong and Lim, Hyungtae and Cho, Younggun},
journal={IEEE Transactions on Field Robotics},
year={2026},
publisher={IEEE}
}