Decentralized Collaborative SLAM

• Decentralized C-SLAM is a method that allows spatially distributed robots to collaboratively map unknown environments without centralized servers.
• It employs factor graphs and distributed optimization techniques to fuse local and inter-robot measurements under communication constraints.
• Empirical evaluations show that decentralized architectures can achieve high-precision mapping with errors in the centimeter to meter range, ensuring scalability and robustness.

Decentralized Collaborative Simultaneous Localization And Mapping (C-SLAM) is the discipline within SLAM that aims to equip teams of spatially distributed robotic agents with the ability to construct a mutually consistent map of an unknown environment and simultaneously localize themselves within it, without recourse to any centralized infrastructure or aggregation server. Agents communicate over peer-to-peer links—often intermittent, lossy, and bandwidth-constrained—sharing compact, partial representations of local state and map features. Core goals include scalability to large teams (e.g., swarms), adaptivity to limited communication, robustness to failures and resource constraints, and high-precision joint localization and mapping in fully unknown or dynamic settings. Over the last decade, the field has evolved from ad hoc fusion of local and inter-robot information to principled, graph- and consensus-based architectures that permit asynchronous operation and strong empirical performance guarantees.

1. Mathematical Formalism and Estimation Objectives

Decentralized C-SLAM generalizes the classical SLAM estimation problem from a single robot to a networked ensemble. The standard estimation objective is maximum a posteriori (MAP) estimation over all robot trajectories and map variables, subject to both local (intra-robot) and shared (inter-robot) measurement constraints. The global MAP cost is typically written as

$$\mathbf{x}^*=\arg\min_{\mathbf{x}} \left[\sum_i \sum_{k} \|h^k_i(x_i)-z^k_i\|_{\Omega^k_i}^2 + \sum_{(i,j)\in E} \sum_{\ell} \|h^\ell_{ij}(x_i,x_j)-z^\ell_{ij}\|^2_{\Omega^\ell_{ij}}\right]$$

where $x_i$ are local state variables of agent $i$, $h^k_i$ denotes the local measurement model, $z^k_i$ the associated measurement, $\Omega$ the information matrix, and $h^\ell_{ij}$ models inter-agent (e.g., loop closure) measurements (Lajoie et al., 2021, Lajoie et al., 2023, McGann et al., 2023, McGann et al., 2024). Each agent need only maintain local factors and the shared factors in which it participates.

To preserve scalability and robustness, state variables and factors are organized into factor graphs—enabling the use of sparse nonlinear least squares or incremental smoothing. In the decentralized setting, no node holds the full graph; instead, robots exchange minimal sufficient subgraphs and/or variable blocks as needed (Lajoie et al., 2023, McGann et al., 2023, McGann et al., 2024, Bird et al., 6 Mar 2025).

2. Algorithmic Architectures and Decentralization Mechanisms

C-SLAM architectures are distinguished by their protocols for state partitioning, factor graph construction, and distributed optimization:

• Local subgraph maintenance: Each agent maintains a pose-graph (or full factor graph) over its own trajectory, local submap, and intra-robot loop closures (Lajoie et al., 2023, Bird et al., 6 Mar 2025, Xu et al., 2022).
• Inter-robot data exchange: Upon rendezvous or network connectivity, robots negotiate what map segments, pose descriptors, and loop closure candidates have not yet been seen and exchange only the missing descriptors/features (Lajoie et al., 2023, McGann et al., 2024).
• Loop closure detection: Inter-robot loop closures may use various global descriptors (e.g., ScanContext, NetVLAD, CosPlace, or foundation-model latents), geometric registration (ICP or TEASER++), or semantic matching (Lajoie et al., 2023, Lajoie et al., 2 Feb 2026, Fernandez-Cortizas et al., 2024).
• Decentralized optimization: Leading approaches include distributed Gauss–Seidel (async or token-based), consensus–ADMM methods (C-ADMM, MESA, iMESA), and asynchronous dual decomposition (ARock) (McGann et al., 2023, McGann et al., 2024, Xu et al., 2022).
• Peer-to-peer synchronization: Protocols are designed to accommodate asynchronous operation, packet loss, and variable topology, often using heartbeat signals, leader election on rendezvous, and versioned message exchanges (Lajoie et al., 2023, Bird et al., 6 Mar 2025, Xu et al., 2022).

The table summarizes some representative architectures:

System	Local Backend	Inter-robot Fusion	Optimization	Sensor Modalities
Swarm-SLAM (Lajoie et al., 2023)	Keyframe Pose-Graph	Global descriptor + geometric verification	Brokered robustified GNC	Lidar, stereo, RGB-D, wheel
$D^2$SLAM (Xu et al., 2022)	Sliding-window VIO	Local bundle adjustment on merged states	ADMM (near), ARock (far)	Omnidirectional/stereo VIO
Multi S-Graphs (Fernandez-Cortizas et al., 2024)	Hierarchical semantic factor graph	Semantic descriptor matching	Local batch (g2o) per robot	Lidar (planar semantics)
iMESA (McGann et al., 2024)	Factor graph/iSAM2	Edge-based splitting w/ consensus ADMM (MESA)	Fully distributed	General (SE(2)/SE(3), Lidar, VIO)

3. Measurement Models, Factor Types, and Data Exchange

C-SLAM exploits a broad array of sensor modalities and associated factor graph elements:

• Monocular/Stereo/RGB-D/VIO: Visual odometry and loop closure via feature correspondences, brute-force descriptor matching, projection factors, and bundle adjustment (Vemprala et al., 2019, Bird et al., 6 Mar 2025, Xu et al., 2022).
• LIDAR: Keyframe-based odometry and place recognition via global descriptors (ScanContext), scan-matching, and ICP (Lajoie et al., 2023, Fernandez-Cortizas et al., 2024, Lajoie et al., 28 Jan 2026).
• Wireless features: Similarity-based loop closure detection via Wi-Fi fingerprints, radio RSSI features, and joint SLAM/crowdsourcing via belief-propagation factor graphs (Yang et al., 2021, Liu et al., 2019).
• Semantic factors: High-level structures such as rooms, planes, and objects are used to compactly encode map constraints, enabling loop closure robust to appearance aliasing and scale (Fernandez-Cortizas et al., 2023, Fernandez-Cortizas et al., 2024).
• Consensus priors and dual variables: In consensus-ADMM formulations, shared variables (poses, map points) are split into local copies with auxiliary variables and duals to enforce consistency (McGann et al., 2023, McGann et al., 2024).

Data exchange is typically dominated by compressed descriptors or factors, not raw point clouds or images. Key bandwidth reduction strategies include budgeted inter-robot edge selection, minimum vertex cover feature transmission, semantic abstraction, and incremental asynchronous keyframe/map-partition transmission (Lajoie et al., 2023, Bird et al., 6 Mar 2025, Fernandez-Cortizas et al., 2024).

4. Distributed and Asynchronous Optimization Techniques

Optimization in C-SLAM must minimize global MAP cost under partitioned data and communication. Main methodologies include:

• Edge-based consensus ADMM (MESA, iMESA): Robots maintain local states plus edge variables for shared (boundary) variables. Each pairwise link enforces a consensus constraint using ADMM, operating directly on the SE(2)/SE(3) manifold. Updates proceed via local nonlinear least squares (e.g., iSAM2), closed-form edge interpolation (e.g., geodesic splitting), and dual ascent steps. iMESA amortizes constraint enforcement over time, yielding near-real-time incremental operation (McGann et al., 2023, McGann et al., 2024).
• Distributed Gauss–Seidel: Each robot cyclically solves its local subproblem (full factor graph with fixed external variables), communicating updated boundary variables (anchor poses) to neighbors. Distributed token-ring or cascaded protocols are used in swarm settings (Niculescu et al., 2024, Lajoie et al., 2023).
• Asynchronous methods (ARock, consensus averaging): Some approaches permit stale or out-of-order updates, relying on eventual consistency as information propagates (Xu et al., 2022).
• Robustified backends: Outlier rejection and robustness are achieved by using GNC (Graduated NonConvexity) loss, pairwise consistency maximization, and information weight scaling by match confidence (Lajoie et al., 2023, Lajoie et al., 2 Feb 2026).
• Incremental/asynchronous keyframe insertion: Systems such as DVM-SLAM use strictly event-driven, idle-cycle processing of externally received map segments, enforcing only local optimization to minimize communication and computational stress (Bird et al., 6 Mar 2025).

Empirical evaluations on standard and real-world datasets consistently show that modern distributed ADMM/Gauss-Seidel pipelines achieve trajectory and map errors within a few percent of centralized approaches, even under significant communication delays and dropouts (McGann et al., 2024, McGann et al., 2023, Bird et al., 6 Mar 2025).

5. Communication, Scalability, and Resource Constraints

C-SLAM is fundamentally constrained by link capacity, agent compute/memory, and network topology.

• Bandwidth budgeting: Swarm-SLAM and related protocols enforce an explicit cap $B$ on inter-robot edge selection per rendezvous. Data per closure ranges from $\sim$4 kB (semantic) to tens of kB (visual/global descriptor), with total session logs on the order of 20–100 MB per robot (Lajoie et al., 2023, Fernandez-Cortizas et al., 2024, Lajoie et al., 28 Jan 2026).
• Scalability: Modern token-ring protocols and budgeted feature transmission enable swarms of up to $O(10^2)$ agents, with per-agent memory budget $<$1.5 MB (nano-UAVs) or $\sim$5–10 MB (industrial) and computational latency for one global optimization round $<$0.3–1 s for $n\sim 10^3$ poses (Niculescu et al., 2024, Lajoie et al., 2023).
• Communication patterns: With event-triggered and on-demand updates, asynchrony, and peer-to-peer synchronization, there is no global clock or obligations on full connectivity. Algorithms accommodate delayed, dropped, or out-of-order packets (McGann et al., 2023, Lajoie et al., 2023).
• Swarm policies: Lightweight C-SLAM stacks may use randomized or state-switching exploration strategies to maximize spatial coverage and minimize collision or local congestion (Niculescu et al., 2024).
• Failure and robustness: Agents tolerate arbitrary local loss or dropouts, as local submaps can later be fused on reconnection. Systems commonly use Mahalanobis-based outlier rejection, dual variable reset, and efficient edge-inactive protocols (Vemprala et al., 2019, McGann et al., 2023).

The following table summarizes resource and scalability regimes:

System	Max Agents	Per-Agent RAM	Mapping RMSE	Bandwidth/run (MB)
Ultra-Light C-SLAM (Niculescu et al., 2024)	200	150 kB	6–30 cm	$\leq$25
Swarm-SLAM (Lajoie et al., 2023)	10–20	100 MB	0.7–5 m	7.5–95 (per exp.)
Multi S-Graphs (Fernandez-Cortizas et al., 2024)	3	5 MB	2–20 cm	0.5–5

6. Sensor, Semantic, and Heterogeneous Integration

C-SLAM frameworks increasingly integrate diverse sensor types and semantic abstraction:

• Visual: Bags-of-Words (DBoW2), NetVLAD, CosPlace, SuperPoint—used for cross-agent place detection and pointwise loop closure (Xu et al., 2022, Bird et al., 6 Mar 2025, Lajoie et al., 2 Feb 2026).
• LIDAR: Global (ScanContext) and local (ICP, TEASER++) registration, planar primitive extraction, and semantic room/wall relational graphs for high-level data reduction and perceptual disambiguation (Lajoie et al., 2023, Fernandez-Cortizas et al., 2024, Fernandez-Cortizas et al., 2023).
• Wireless, Acoustic: Radio fingerprinting, AOA/TOA/BP-based SLAM, and crowdsourced feature maps, allowing decentralized mapping even in GPS-denied, visually ambiguous, or infrastructure-poor settings (Liu et al., 2019, Yang et al., 2021).
• 3D foundation models: Used for robust loop closing across extreme viewpoint change, leveraging foundation model latents and confidence-weighted outlier suppression (Lajoie et al., 2 Feb 2026).
• Heterogeneous agents: Decentralized SLAM supports teams with mixed sensing and estimation backends (visual, LIDAR, semantic), fusing only overlapping marginal beliefs or factor blocks as needed (Dagan et al., 2023).

7. Empirical Evaluation, Robustness, and Future Directions

Extensive quantitative and qualitative evaluations have established the effectiveness of decentralized C-SLAM architectures:

• Typical errors for pose (ATE) and map are within 2–30 cm (lightweight systems (Niculescu et al., 2024)), 1.1–4.5 cm (dense LIDAR, visual), or below 1 m (field-scale GPS-free) in controlled and real-world environments up to 475 m trajectories (Lajoie et al., 2023, Lajoie et al., 28 Jan 2026, Bird et al., 6 Mar 2025, Fernandez-Cortizas et al., 2024).
• Robust estimation is maintained under severe resource constraints, perceptual aliasing, delayed communication, and partial network partitioning through robustified solvers and semantic abstraction (Lajoie et al., 2023, Lajoie et al., 2 Feb 2026, Fernandez-Cortizas et al., 2024).
• Ongoing research targets include: integration of event-triggered communication, consensus-based high-level semantic mapping, distributed foundation-model place recognition, formal convergence and optimality guarantees for nonconvex distributed optimization, and scalable support for heterogeneous, dynamic robot collectives (McGann et al., 2023, McGann et al., 2024, Fernandez-Cortizas et al., 2023, Dagan et al., 2023).

Decentralized C-SLAM stands as a cornerstone for the next generation of resilient, scalable, multi-agent robotics in environments ranging from Mars-analogue deserts to ultra-dense indoor swarms, enabling collaborative environmental inference without reliance on centralized infrastructure or constant connectivity (Lajoie et al., 28 Jan 2026, Niculescu et al., 2024, Xu et al., 2022, Lajoie et al., 2021, Lajoie et al., 2023).