Decentralized Collaborative SLAM

• Decentralized Collaborative SLAM is a framework enabling multiple autonomous agents to locally estimate poses and collaboratively build a global map without a central server.
• It employs distributed optimization techniques like pose-graph optimization and consensus-based methods to ensure robust, scalable, and globally consistent mapping under communication constraints.
• The approach leverages semantic representations and adaptive data fusion protocols to reduce bandwidth usage and enhance resilience in diverse, resource-constrained environments.

Decentralized Collaborative Simultaneous Localization and Mapping (C-SLAM) refers to a suite of algorithms, system architectures, and protocols enabling multiple autonomous agents to jointly estimate their own poses and construct a geometric or semantic map of an unknown environment, without reliance on a central server or persistent infrastructure. In this paradigm, each agent runs local perception, estimation, and mapping processes; communicates selectively with peers to exchange measurements, descriptors, or summarized map representations; and participates in peer-to-peer (or network-based) optimization or data fusion to achieve global spatial consistency, efficient resource utilization, and robustness to failures or communication disruptions.

1. Key Principles and Formal Problem Statement

Let a robotic team of $N$ agents, indexed by $i=1\dots N$, operate in a bounded or unbounded domain. Each robot $i$ possesses its own state trajectory $\mathcal{X}_i = \{x_i^1, \dots, x_i^{T_i}\}$, where each $x_i^t \in SE(3)$ or $SE(2)$ denotes the robot's pose at time $t$. Each robot makes:

• Intra-robot measurements $z_i^\ell$ (e.g., odometry, local loop closures)
• Inter-robot measurements $z_{ij}^k$, relating $x_i^{t_i}$ to $x_j^{t_j}$ (e.g., via shared landmarks or place recognition)

The global collaborative estimation objective is to maximize the joint posterior:

$$\{\mathcal{X}_i^*\} = \arg\max_{ \{ \mathcal{X}_i \} } p(\mathcal{X}_1, ..., \mathcal{X}_N | \mathcal{Z})$$

or equivalently in nonlinear least-squares form under Gaussian measurement noise (Lajoie et al., 2021):

$$\{\mathcal{X}_i^*\} = \arg\min_{\{X_i\}} \sum_{i=1}^N \sum_\ell \| h_i^\ell(X_i^\cdot) - z_i^\ell \|^2_{\Omega_i^\ell} + \sum_{i<j} \sum_k \| h_{ij}^k(x_i^{t_i}, x_j^{t_j}) - z_{ij}^k \|^2_{\Omega_{ij}^k}$$

The challenge of decentralized C-SLAM lies in realizing this inference reliably and efficiently, where each agent holds only partial information, communication is constrained, and system-wide consensus must emerge via distributed or opportunistic data fusion.

2. Core Algorithmic Frameworks

Decentralized C-SLAM admits multiple architectural taxonomies (Lajoie et al., 2021), including:

Approach	Representative Methods	Data Exchanged
Distributed Pose-Graph Optimization	Block Gauss-Seidel, DGS (McGann et al., 2023)	Spanning sets of local updates
Consensus-Based Estimation	On-manifold ADMM, MESA (McGann et al., 2023)	Averaged state estimates
Submap Merging and Rendezvous-Based	Hybrid, Spectral Sparsification	Submap descriptors/constraints
Factor-Graph Partitioning	Multi-root iSAM, iMESA (McGann et al., 2024)	Separator clique messages

Distributed pose-graph optimization decomposes the factor graph by robot ownership; each agent holds its local subgraph and iteratively updates states treating neighbor estimates as fixed, exchanging only relevant updates. Consensus-based approaches exploit variants of ADMM or distributed averaging on the manifold (e.g., SE(3)), allowing asynchronous edge-based updates without full synchronization (McGann et al., 2023, McGann et al., 2024). Submap merging leverages local batch maps or semantic substructures, aligning these at rendezvous events using geometric or semantic place recognition. Factor-graph partitioning enables incremental solvers (e.g., iSAM2) to operate in parallel, sharing marginals over separator variables (McGann et al., 2024, Dagan et al., 2023).

Manifold-compatible ADMM or consensus algorithms, such as MESA, introduce variables $z_{(i,j),s}$ and dual multipliers $\lambda_{(i,j),s}$ for each shared state $s$, minimizing the augmented Lagrangian with manifold difference functions and dual penalties (McGann et al., 2023).

3. Semantic-Relational and High-Level Representations

To address the scalability and robustness limits of raw feature-based CSLAM, hierarchical semantic-relational graphs and object-based descriptors have emerged (Fernandez-Cortizas et al., 2024, Fernandez-Cortizas et al., 2023, Liu et al., 2024). Multi S-Graphs, for instance, encode environments as a four-layered hierarchical S-Graph:

• Keyframe Layer: $X_k \in SE(3)$ for each LiDAR frame with odometry constraints $f_{ij}$
• Wall Layer: plane nodes $P_i$ parameterized as $\mathbb{R}^3$ (normals, distances)
• Room Layer: nodes $R_m$ for semantic rooms/corridors, connected to walls
• Floor Layer: a global node $F$ per building floor

Inter-robot communication and loop closure are mediated via compact, semantic descriptors—e.g., a “room descriptor” generated by downsampling the composite room point cloud and extracting a ScanContext matrix. Only upon semantic match and geometric validation (e.g., VGICP) are high-fidelity point clouds or registration factors exchanged. This achieves 90–97% bandwidth reduction compared to raw feature exchange, with robust avoidance of false loop closures (Fernandez-Cortizas et al., 2024).

Similarly, SlideSLAM exploits an object-level sparse semantic map (cuboids, cylinders, ellipsoids) and a hypothesis-driven loop-closure strategy, achieving sub-decimeter localization with bandwidth on the order of 1.65–5.13 KB/m mapped, and inter-robot pose errors of 0.22 m ± 0.15 m (Liu et al., 2024).

4. Communication and Data Association Protocols

Bandwidth-efficient and adaptive communication is central to decentralized C-SLAM. Common principles include:

• Hierarchical Data Distillation: Exchange only topological/semantic graphs or minimal signatures, with on-demand retrieval of raw data for validation (Fernandez-Cortizas et al., 2024).
• Peer-to-peer Opportunism: Neighbors communicate over ad hoc wireless, exchanging descriptors, candidate loops, or small-factor cliques only when in range or scheduled (Lajoie et al., 2023, Fernandez-Cortizas et al., 2024).
• Prioritization and Sparsification: Prioritized selection of loop-closure candidates via spectral connectivity, maximizing algebraic connectivity of the joint pose graph under a communication budget (Lajoie et al., 2023).
• Consistency Maintenance: Brokers or anchor election maintain a consistent global reference; serverless rendezvous-based protocols ensure asymptotic agreement with only partial peer-to-peer connectivity (Lajoie et al., 2023, Lajoie et al., 28 Jan 2026).
• Multi-modal Fusion: Abstract all edges (odometry, loop closures, cross-modal) as $SE(3)$ constraints to enable universal back-end optimization (Lajoie et al., 2023).
• Plug-and-play Data Fusion: For heterogeneous teams, only relevant marginal factors are exchanged along conditional independence boundaries (e.g., target states in multi-target SLAM+tracking), maintaining statistical consistency via channel filters (Dagan et al., 2023).

5. Distributed Back-end Optimization and Incremental Solvers

Modern decentralized C-SLAM systems employ distributed nonlinear optimization frameworks, frequently built atop incremental solvers (iSAM2) with distributed consensus extensions:

• Separable Manifold ADMM (MESA/iMESA): Each agent’s state is optimized locally, with consensus constraints over shared variables enforced via dual variables and edge-averaging (spherical interpolation for rotations) (McGann et al., 2023, McGann et al., 2024). iMESA offers a fully incremental, event-driven protocol: local updates are triggered by new measurements, while pairwise communication reconciles shared states and activates ADMM penalties.
• Edge-based Asynchrony: Only neighbor pairs with communication perform updates, allowing robust operation under packet loss and time-varying topology (McGann et al., 2023).
• Scalability: Communication cost per iteration remains proportional to the number of shared variables, enabling scaling to large multi-robot teams with >10,000 state variables (McGann et al., 2024).
• Convergence: Empirical results show convergence to centralized accuracy within a small multiple of local update iterations, generally outperforming prior DDF-SAM or naive consensus algorithms under sparse communication (McGann et al., 2024).

6. Experimental Results, Performance, and Applications

Recent works validate decentralized C-SLAM across diverse agent types, sensing modalities, and environments:

• Structured Indoor Environments: Multi S-Graphs achieves 40% lower mapping time than single-agent, near-zero false positives in highly symmetric corridors, and two orders of magnitude data reduction per agent (Fernandez-Cortizas et al., 2024, Fernandez-Cortizas et al., 2023).
• Swarm and Low-resource Agents: Ultra-Lightweight Collaborative Mapping supports centimeter-level mapping in swarms of up to 100 nano-UAVs, each with <1.5 MB RAM, by employing token-based scan exchange and distributed Gauss–Seidel optimization (Niculescu et al., 2024).
• Heterogeneous Teams: SlideSLAM demonstrates accurate, viewpoint-robust semantic mapping and <25 cm pose drift in multi-modal, multi-agent outdoor/indoor scenarios (LiDAR/RGB-D/UAV/UGV) (Liu et al., 2024).
• Planetary Analogue and Communication Constrained Regimes: Robustness to sparse features, high latency (>100 ms), and intermittent connectivity is achieved by dynamic communication budgeting, loop-closure prioritization, and local outlier rejection (Lajoie et al., 28 Jan 2026).
• Aerial Swarms: $D^2$SLAM offers centimeter-level ego-motion error and global consistency under limited field-of-view and bandwidth, using distributed optimization (ADMM, ARock) and dual camera front-ends (Xu et al., 2022).
• Plug-and-Play Crowdsourcing: Radio-based agents achieve decimeter-level localization in multipath-challenged settings by decoupling measurement biases and merging local feature maps with probabilistic existence weights (Yang et al., 2021).

7. Limitations, Open Problems, and Future Directions

While decentralized C-SLAM systems now demonstrate robust, scalable, and accurate mapping in diverse operational settings, several open challenges persist:

• Semantic Descriptor Generalization: Approaches relying on semantic or geometric distinctiveness (e.g., room geometry, sparse objects) can degrade in alias-prone or open-plan spaces (Fernandez-Cortizas et al., 2024, Liu et al., 2024). Learning-based or hierarchical descriptors remain a critical research direction.
• Scalability to Large Agent Teams: While evidence shows sublinear growth in communication and computation, handling hundreds to thousands of agents (e.g., nanodrone swarms) necessitates advanced traffic shaping, dynamic tokenization, and hierarchical optimization (Niculescu et al., 2024).
• Heterogeneous Algorithm Integration: Ensuring statistical consistency and architectural interoperability across disparate SLAM engines (metric, semantic, LiDAR, vision) is addressed via message marginalization and black-box fusion (Dagan et al., 2023), but full theoretical guarantees under arbitrary algorithmic heterogeneity remain open.
• Resiliency to Adversity: High-vibration, dust, poor-features, and non-line-of-sight scenarios (e.g., planetary terrain) expose limits in local odometry and loop-closure reliability, requiring data-driven parameter adaptation, robust filtering, and multi-sensor fusion (Lajoie et al., 28 Jan 2026).
• Global Consistency and Feedback: While most current frameworks maintain local or pairwise consistency, achieving temporally-global optimality (across all agents and time) under minimal feedback is still an area of methodological innovation (Fernandez-Cortizas et al., 2023).

In summary, Decentralized Collaborative SLAM represents a rapidly advancing frontier in multi-robot perception and autonomy, integrating semantic abstraction, distributed optimization, and bandwidth-aware communication to deliver scalable, robust, and real-time mapping capabilities in resource-constrained, unpredictable environments (Fernandez-Cortizas et al., 2024, Fernandez-Cortizas et al., 2023, McGann et al., 2023, McGann et al., 2024, Niculescu et al., 2024, Lajoie et al., 2023, Liu et al., 2024, Xu et al., 2022, Bird et al., 6 Mar 2025, Lajoie et al., 28 Jan 2026).