DRACo-SLAM2: Efficient Underwater SLAM

• The paper introduces a fully distributed SLAM framework that replaces costly global scan-matching with lightweight object-graph representations for enhanced underwater communication.
• It employs an eigenvalue-based graph matching pipeline and an incremental group-wise consistent measurement set maximization to robustly reject outliers in inter-robot loop closures.
• Empirical results show a 28× reduction in ICP time and a 10–20% improvement in trajectory accuracy, validated on both simulated and real-world deployments.

Searching arXiv for the specified paper and closely related items to ground the article in the current record. arxiv_search(query="(Huang et al., 31 Jul 2025) OR DRACo-SLAM2 OR Distributed Robust Acoustic Communication-efficient SLAM for Imaging Sonar Equipped Underwater Robot Teams with Object Graph Matching", max_results=5, sort_by="relevance") DRACo-SLAM2 is a fully distributed, communication-efficient SLAM framework for teams of underwater robots equipped with imaging sonar. It extends the original DRACo-SLAM by replacing costly global scan-matching with a lightweight object-graph representation of the sonar map, an eigenvalue-based graph matching pipeline for initial alignment, and an incremental Group-wise Consistent Measurement Set Maximization (GCM) procedure for robust rejection of outliers in inter-robot loop closures. The framework targets underwater acoustic networks, where communication is severely bandwidth-limited, and it is validated on simulated and real-world datasets with comparative analyses against DRACo-SLAM1 and alternative optimization choices (Huang et al., 31 Jul 2025).

1. System architecture and operating regime

Each robot $\alpha$ maintains three local structures: a sonar-only SLAM back end that ingests vehicle dead reckoning and sequential-scan match measurements to produce a pose graph, a local $2$D point-cloud map $M_\alpha$ formed by transforming salient sonar returns into the robot frame, and a derived object map $V_\alpha$ obtained by clustering $M_\alpha$ with DBSCAN and fitting axis-aligned bounding boxes (Huang et al., 31 Jul 2025). In the reported implementation, the local SLAM back end uses Bruce-SLAM [20].

When two robots $\alpha$ and $\beta$ come within acoustic communication range, with typical bandwidth $\approx 62.5$ kbps, they exchange their current object maps $V_\alpha$ and $V_\beta$ rather than full point clouds or high-dimensional descriptors. A bipartite graph matching module computes an initial rigid alignment $2$0 between the two object graphs. Robot $2$1 then requests only the minimal set of full sonar scans needed to perform a local ICP registration, using $2$2 as the starting guess. Inter-robot loop closures that survive a robust outlier filter are inserted as constraints into a two-step inter-robot pose-graph optimizer.

Imaging sonar has two stated roles in this architecture. First, it supplies per-scan point clouds for local mapping and ICP registration. Second, it provides raw returns from which objects are extracted and abstracted into compact graph vertices. The framework is therefore not a graph-only system: the object graph is used to reduce search and communication costs, while geometric refinement remains scan-based. This structure directly addresses a common misunderstanding that DRACo-SLAM2 eliminates scan registration; in fact, it defers and localizes scan exchange and ICP rather than removing them.

2. Local estimation, object extraction, and graph construction

The local estimation problem is posed in standard factor-graph form as a MAP objective,

$2$3

where $2$4 are $2$5DOF poses, $2$6 are dead-reckoning odometry, $2$7 are sequential-scan matches, and $2$8 are intra-robot loop closures (Huang et al., 31 Jul 2025).

From each sonar image at time $2$9, feature points $M_\alpha$0 are detected via a CFAR-based method and accumulated into a local point cloud,

$M_\alpha$1

The latest $M_\alpha$2 is clustered with DBSCAN. Each cluster yields an object $M_\alpha$3 described by its centroid $M_\alpha$4 and bounding-box dimensions $M_\alpha$5. After discarding clusters with too few points or small extent, the remaining objects define the object map

$M_\alpha$6

The object map is then converted into a complete directed graph

$M_\alpha$7

with edge weights

$M_\alpha$8

This representation compresses the sonar map into vertices carrying centroid and axis-aligned extent information and edges carrying pairwise Euclidean distances. A plausible implication is that the representation preserves enough relational structure for inter-robot association while remaining compact enough for routine acoustic transmission. The reported communication analysis is consistent with that interpretation.

3. Object graph matching and local scan registration

Given two object graphs $M_\alpha$9 and $V_\alpha$0, DRACo-SLAM2 seeks a partial bijection between $V_\alpha$1 and $V_\alpha$2 that maximizes structural similarity. The matching problem is formulated as a Quadratic Assignment Problem,

$V_\alpha$3

where $V_\alpha$4 is a binary correspondence matrix subject to $V_\alpha$5 and $V_\alpha$6 (Huang et al., 31 Jul 2025).

The pairwise-edge utility is defined as

$V_\alpha$7

By stacking $V_\alpha$8, the problem is recast as an eigen-problem,

$V_\alpha$9

where $M_\alpha$0 encodes all $M_\alpha$1. Taking the principal eigenvector as $M_\alpha$2 and then solving

$M_\alpha$3

via the Jonker-Volgenant algorithm yields the discrete correspondence $M_\alpha$4. From at least three matched object pairs, an initial rigid transform $M_\alpha$5 is estimated by RANSAC-filtered affine fitting.

The transform from graph matching is used only as an initial alignment. Robot $M_\alpha$6 requests from $M_\alpha$7 the small subset of raw sonar scans whose extracted feature points overlap the matched objects, and a local ICP with a sliding window of size $M_\alpha$8 frames refines the alignment. The paper states that this graph-based prior brings the two clouds within ICP’s basin of convergence and that, compared to the Go-ICP used in DRACo-SLAM1, the two-stage process is $M_\alpha$9 faster on average (Huang et al., 31 Jul 2025).

4. Incremental Group-wise Consistent Measurement Set Maximization

Outlier rejection is handled by incremental Group-wise Consistent Measurement Set Maximization. The motivation is that nearby or similar scan pairs can share correlated registration biases. Standard Pairwise Consistent Measurement Maximization (PCM) checks only pairwise consistency within a single inter-robot edge set $\alpha$0:

$\alpha$1

However, when $\alpha$2 has loop closures to multiple neighbors such as $\alpha$3 and $\alpha$4, an outlier in the $\alpha$5-$\alpha$6 set may remain mutually consistent with other $\alpha$7-$\alpha$8 edges while still being systematically biased (Huang et al., 31 Jul 2025).

GCM extends the consistency check across the entire local group by evaluating, for example,

$\alpha$9

with

$\beta$0

The $\beta$1 terms are marginalized relative-pose estimates from local SLAM or previously accepted loop closures. The method then seeks the largest subset of loop closures $\beta$2 for which all pairwise $\beta$3, solvable by a simple greedy or maximum-clique extraction.

The reported empirical effect is specific: GCM rejects groups of biased but mutually consistent outliers that PCM alone would admit. This is the central methodological distinction between DRACo-SLAM2 and pairwise-only consistency screening.

5. Two-step inter-robot pose-graph optimization and communication model

For a robot set $\beta$4, the full multi-robot MAP problem is written as

$\beta$5

with $\beta$6 and $\beta$7 (Huang et al., 31 Jul 2025). Rather than solving this monolithically, DRACo-SLAM2 uses a two-step procedure.

In Step 1, a local “star-shaped” optimization is performed around $\beta$8:

$\beta$9

where $\approx 62.5$0. This updates $\approx 62.5$1 and only those poses of its neighbors involved in $\approx 62.5$2’s current loop closures. In Step 2, each neighbor $\approx 62.5$3 is locally re-optimized as

$\approx 62.5$4

using $\approx 62.5$5’s prior marginalized SLAM poses $\approx 62.5$6 and the historical subset $\approx 62.5$7 from Step 1. All inter-robot constraints enter the graph with a robust Cauchy noise model.

The communication model follows the same economy. At each sonar time step, a robot broadcasts the current object map $\approx 62.5$8 and incremental updates to its marginalized pose history. Only upon successful graph matching does it request raw sonar scans linked to matched objects. The reported message sizes and bandwidth comparisons are summarized below (Huang et al., 31 Jul 2025).

Quantity	Reported value
Typical acoustic bandwidth	$\approx 62.5$9 kbps
DRACo-SLAM1 ring-key descriptor	$V_\alpha$0 Kbits
Full submaps per contact	$V_\alpha$1 Kbits
DRACo-SLAM2 object map, average	$V_\alpha$2 Kbits
DRACo-SLAM2 object map, maximum	$V_\alpha$3 Kbits
Continuous overhead reduction	$V_\alpha$4

The stated interpretation is that object-map messages remain well below typical acoustic limits, while point-cloud exchanges remain identical to DRACo1. A common misconception would be to treat DRACo-SLAM2 as purely descriptor-based; instead, it uses sparse graph exchange continuously and raw scan exchange selectively.

6. Experimental results, limitations, and research directions

The framework is validated on two fully simulated $V_\alpha$5-robot scenarios, the “Airplane” and “USMMA” datasets in HoloOcean, and on a real USMMA deployment (Huang et al., 31 Jul 2025). For inter-robot loop closure detection, the reported precision and true-positive loop-closure count are studied as functions of the minimum overlap ratio threshold $V_\alpha$6. With a sliding window of $V_\alpha$7 frames, DRACo-SLAM2 maintains $V_\alpha$8 precision while tripling the number of detected true positives compared to the global-ICP-based DRACo1. The mean ICP time falls from $V_\alpha$9 ms to $V_\beta$0 ms, corresponding to an $V_\beta$1 speed-up, and the average per-timestep registration calls rise from $V_\beta$2 to $V_\beta$3, enabling more frequent and timely loop closures.

For inter-robot pose-graph optimization, Absolute Trajectory Error RMSE in meters is compared for the two-step PGO with PCM and with GCM against a full monolithic PGO. DRACo-SLAM2 with GCM yields the lowest ATE in $V_\beta$4 of $V_\beta$5 trajectories and typically improves over PCM by $V_\beta$6–$V_\beta$7. Although the two-step PGO slightly decouples remote estimates, it runs $V_\beta$8–$V_\beta$9 faster in large teams. In the real-world USMMA deployment, fused trajectories from three BlueROV2s are shown aligned with satellite imagery, and communication logs confirm that object-map messages of $2$00 Kbits remain an order of magnitude below modem capacity while yielding robust global alignment without drift.

Evaluation item	Reported result
Precision with sliding window of 3 frames	$2$01
True positives vs. DRACo1	Tripled
Mean ICP time	$2$02 ms $2$03 ms
Average registration calls per timestep	$2$04
Lowest ATE trajectories with GCM	$2$05 of $2$06
ATE improvement of GCM over PCM	$2$07–$2$08
Two-step PGO speed in large teams	$2$09–$2$10 faster

The paper also identifies explicit limitations. Repetitive or highly symmetric structures, such as parallel pilings, can still induce perceptual aliasing in graph matching; tighter semantic cues or sonar-intensity-based object descriptors may help. The current object-map update policy shares the full latest map each timestep, and a delta-based protocol could further cut bandwidth. Dynamic obstacles such as fish or swimmers may corrupt DBSCAN clusters, leaving temporal filtering or sonar-based motion segmentation as open questions. Extending GCM beyond the star-center $2$11 setting to fully exploit multi-robot cycle consistency, or incorporating covariance-aware consistency metrics, is described as a promising research avenue. The framework is also positioned for coupling with active exploration planners to support fully autonomous, long-duration surveys with minimal human supervision.

Taken together, these results define DRACo-SLAM2 as a sonar-centric distributed SLAM system in which sparse object graphs, local scan verification, and group-wise consistency filtering are jointly used to deliver robust loop closure detection and real-time multi-robot performance under underwater acoustic communication constraints (Huang et al., 31 Jul 2025).