Distributed Pose Graph Optimization

• Distributed Pose Graph Optimization is a decentralized approach for solving nonlinear pose estimation in multi-robot SLAM by partitioning the global graph into local subproblems.
• It employs various decomposition paradigms, such as block-iterative, proximal, and consensus methods, to address scalability, robustness, and communication constraints.
• Key algorithmic families include ADMM, low-rank Riemannian, and learned policies, each offering practical trade-offs between accuracy and computational efficiency.

Distributed Pose Graph Optimization (PGO) denotes the solution of pose-graph optimization under decentralized, distributed, or parallel computational constraints, typically in multi-robot SLAM, collaborative localization, camera-network localization, and related networked estimation settings. In its standard form, the underlying problem is a nonlinear maximum-likelihood estimation over poses in $SE(d)$, with relative measurements on graph edges and unknown absolute poses at graph vertices. The distributed variant preserves the pose-graph objective while replacing a centralized solver by local computation on graph partitions, robot-owned subgraphs, or consensus-coupled variable blocks, with only neighbor-to-neighbor communication or other restricted message passing. Across the literature, distributed PGO is described through several closely related lenses: nonlinear least-squares on $SE(d)$, $SE(d)$-synchronization, low-rank semidefinite relaxation, Riemannian optimization on product manifolds, block coordinate or proximal majorization schemes, consensus dynamics, ADMM-based splitting, and asynchronous parallel optimization (Choudhary et al., 2017, Tian et al., 2019, Li et al., 2024, Sonawalla et al., 3 Mar 2026).

1. Problem statement and mathematical structure

The canonical distributed PGO problem inherits the centralized pose-graph formulation. Each node $i$ represents an unknown pose $T_i=(R_i,t_i)\in SE(d)$, with $R_i\in SO(d)$ and $t_i\in\mathbb R^d$, and each edge $(i,j)$ provides a noisy relative measurement $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$. A widely used maximum-likelihood objective is

$$\min_{R_i\in SO(d),\,t_i\in\mathbb R^d} \sum_{(i,j)\in E} \kappa_{ij}\|R_j-R_i\widetilde R_{ij}\|_F^2 + \tau_{ij}\|t_j-t_i-R_i\widetilde t_{ij}\|_2^2,$$

or equivalent geodesic variants based on $SE(d)$0 (Tian et al., 2019, Tian et al., 2020, Li et al., 2024, Sonawalla et al., 3 Mar 2026). Other formulations specialize to planar multi-robot systems with relative range and odometry factors, as in DGORL, where robot $SE(d)$1 at time $SE(d)$2 carries $SE(d)$3 and the cost combines range residuals and motion-prior residuals over a sliding window (Latif et al., 2022).

Two features distinguish the distributed setting. First, the global variable set is partitioned by robot, subgraph, or ownership rule, so that each agent optimizes only a subset of poses or a local copy of shared variables. Second, the global cost is rewritten as a sum of local objectives plus coupling terms induced by inter-robot loop closures, separator poses, duplicated object poses, or overlap regions (Choudhary et al., 2017, Tian et al., 2020, Zhao et al., 23 Jun 2026). This decomposition preserves sparsity: robot $SE(d)$4 needs only the neighboring states or boundary variables that appear in factors touching its local variables.

Several papers interpret distributed PGO as a synchronization problem. In the multi-level partitioning framework of Li et al., the global nonlinear objective is explicitly cast as $SE(d)$5-synchronization and then decomposed after graph partitioning into balanced local objectives $SE(d)$6 over subsets $SE(d)$7 (Li et al., 2024). Other work adopts lifted or factorized forms. Distributed certifiably correct PGO introduces a sparse semidefinite relaxation over a positive semidefinite matrix $SE(d)$8, followed by a Burer–Monteiro factorization $SE(d)$9 with $SE(d)$0 (Tian et al., 2019). ASAPP uses a rank-restricted relaxation over Stiefel variables $SE(d)$1 and Euclidean variables $SE(d)$2, then groups variables per robot to obtain

$SE(d)$3

where $SE(d)$4 (Tian et al., 2020).

This common structure makes distributed PGO a family of nonconvex, sparse, graph-coupled optimization problems on manifolds. A plausible implication is that many apparently different methods differ less in objective than in their treatment of coupling, local subproblem geometry, and communication discipline.

2. Decomposition paradigms and communication models

A central design choice is how the global pose graph is split across agents. Early distributed mapping work partitions variables by robot trajectory and identifies “separator” variables as the poses appearing in inter-robot factors. The resulting global Hessian and gradient decompose blockwise into $SE(d)$5, $SE(d)$6, and $SE(d)$7, with $SE(d)$8 only when robots $SE(d)$9 and $i$0 are linked by measurements (Choudhary et al., 2017). This robot-native partition is simple, but later work emphasizes that it can be highly unbalanced when robots generate unequal numbers of keyframes.

Balanced redistribution is the defining idea of BDPGO. Instead of fixing subproblems to native robot trajectories, BDPGO redistributes poses across the swarm through a two-stage graph-partitioning procedure: a streaming FENNEL-style assignment followed by periodic ParMETIS-based repartitioning subject to near-uniform cardinality constraints. The stated purpose is to reduce subproblem imbalance, decrease communication volume, and preserve coherence under robot failure or network changes (Xu et al., 2021). Multi-level partitioning pursues a related objective in collaborative SLAM. Li et al. employ a classical coarsening-initial-partitioning-uncoarsening pipeline, evaluate four KaHIP configurations, and report that the “Highest” scheme produces the lowest cut-edge counts and inter-subgraph correlation among the tested settings (Li et al., 2024).

A different decomposition principle is overlap. ROBO defines for robot $i$1 an overlap set

$i$2

so each robot optimizes not only its owned poses but also a graph-distance neighborhood. The local subproblem includes interior edges and boundary terms involving fixed neighbor poses. At $i$3 the method reduces to disjoint domains, whereas $i$4 approaches centralized optimization (Sonawalla et al., 3 Mar 2026). This provides a continuum between fully distributed and centralized regimes, with the overlap radius acting as an explicit communication–convergence control parameter.

Communication assumptions vary, but most methods use local message passing on either the SLAM graph or a robot interaction graph. GeoD is distributed over the pose graph itself, with a separate computation thread for each node and messages passed only between neighboring nodes (Cristofalo et al., 2020). ASAPP adopts a two-thread architecture per robot: a communication thread that maintains cached neighbor variables and an optimization thread driven by a Poisson clock (Tian et al., 2020). DRAN explicitly decouples the communication topology from physical interaction topology by imposing consensus only over a communication graph $i$5, which may be sparse, intermittent, or time-varying (Zhao et al., 23 Jun 2026).

The literature repeatedly contrasts local transmission of variables with centralized exchange of raw sensor data. In the distributed mapping framework of Choudhary et al., each iteration communicates only separator estimates, and the use of object-based models further avoids sharing raw point clouds or imagery (Choudhary et al., 2017). This emphasis on sparse state exchange, rather than raw measurement exchange, is one of the defining operational features of distributed PGO.

3. Major algorithmic families

Distributed PGO has developed along several major algorithmic lines rather than a single dominant solver class.

The first major family comprises linearized block-iterative methods derived from two-stage relaxations. Choudhary et al. propose distributed implementations of the two-stage approach of Carlone et al., using Jacobi Over-Relaxation and Successive Over-Relaxation; the latter with $i$6 is the Distributed Gauss-Seidel method or DGS. In this framework, robots alternately solve local linear systems while exchanging only separator-variable estimates, and DGS is reported to have excellent empirical performance and an anytime flavor (Choudhary et al., 2017). DGORL is also block-iterative, but it is built atop distributed g$i$7o optimization over dynamic communication graphs with damped Gauss–Newton or Levenberg–Marquardt updates and weight-averaged neighbor mixing (Latif et al., 2022).

The second family consists of proximal, majorization, and coordinate-descent methods on manifolds. The generalized proximal framework writes the PGO cost as $i$8 and constructs a block-separable surrogate using an upper-bounding matrix $i$9, yielding per-node closed-form updates involving one $T_i=(R_i,t_i)\in SE(d)$0 SVD and translation correction (Fan et al., 2020). Closely related MM methods build quadratic surrogates or majorizers around the current iterate and then solve decomposed local subproblems in parallel. Both unaccelerated MM-PGO and accelerated AMM-PGO are proved to converge to first-order critical points under mild conditions, and the framework extends to robust losses such as Huber and Welsch (Fan et al., 2020, Fan et al., 2021). Li et al. develop an Improved Riemannian Block Coordinate Descent (IRBCD) algorithm on a low-rank manifold factorization, with momentum mixing, block updates, adaptive restart, and a convergence argument to a globally optimal critical point under rank sufficiency (Li et al., 2024).

The third family is consensus-based dynamics. GeoD minimizes a geodesic pose-graph error through continuous-time distributed consensus laws on rotation and translation: $T_i=(R_i,t_i)\in SE(d)$1 Lyapunov analysis, pairwise consistency, and minimal consistency conditions are used to prove convergence to equilibria corresponding to local minima, after a one-step distributed initialization round (Cristofalo et al., 2020). CORD generalizes this dynamical viewpoint by treating PGO as a second-order damped Euler–Poincaré system on $T_i=(R_i,t_i)\in SE(d)$2, with mass and damping matrices and a semi-implicit geometric integrator. In the overdamped limit, the resulting update recovers Riemannian gradient descent, while suitable damping choices recover Gauss–Newton-like behavior (Shin et al., 11 May 2026).

The fourth family is ADMM and splitting. The SOC–ADMM/Bregman method introduces auxiliary matrices $T_i=(R_i,t_i)\in SE(d)$3 and scaled duals $T_i=(R_i,t_i)\in SE(d)$4, alternates unconstrained quadratic rotation updates with orthogonality-constrained Procrustes projections $T_i=(R_i,t_i)\in SE(d)$5, and then solves translations as a quadratic system (Ebrahimi et al., 10 Mar 2025). PieADMM approaches 3D PGO through unit quaternions and translations under a von Mises–Fisher rotation model, uses manifold projections by normalization, and derives a proximal linearized Riemannian ADMM with parallelizable local subproblems and $T_i=(R_i,t_i)\in SE(d)$6 iteration complexity for $T_i=(R_i,t_i)\in SE(d)$7-stationarity (Chen et al., 2024).

The fifth family is certifiable low-rank Riemannian optimization. Distributed certifiably correct PGO starts from a sparse SDP relaxation, solves its low-rank factorization by Riemannian block coordinate descent, and then performs distributed solution verification via eigenvalue computations on a dual certificate matrix $T_i=(R_i,t_i)\in SE(d)$8. If the certificate is negative, a saddle-escape direction is constructed by rank lifting, after which the algorithm restarts at higher rank (Tian et al., 2019). This is distinct from first-order convergence guarantees: it aims at distributed recovery of globally optimal solutions under moderate measurement noise, matching centralized certifiable methods.

A sixth line departs from classical optimization and uses learned policies. “Policies over Poses” casts planar distributed PGO as a partially observable Markov game over partitioned local pose graphs. Each agent uses a recurrent edge-conditioned GNN encoder, adaptive edge-gating denoiser, GRU memory, and a hybrid actor that chooses an edge and a small pose correction; an information-weighted asynchronous ADMM reconciles separator poses afterward (Ghanta et al., 26 Oct 2025). This suggests that distributed PGO has become a platform for both analytical and learning-based solvers rather than a narrowly defined numerical subfield.

4. Initialization, consistency, robustness, and optimality certification

Initialization is repeatedly treated as decisive because distributed PGO remains nonconvex. Several methods rely on chordal, spectral, or relaxed initializers. ASAPP uses distributed chordal initialization plus 50 Gauss–Seidel steps on benchmark datasets (Tian et al., 2020). The generalized proximal method reports strong performance from chordal initialization on both 2D and 3D SLAM benchmarks (Fan et al., 2020). Majorization-based methods also include accelerated distributed chordal initialization with $T_i=(R_i,t_i)\in SE(d)$9 convergence for the convex rotation-relaxation stage, followed by projection onto $R_i\in SO(d)$0 and translation recovery (Fan et al., 2020).

GeoD takes a different route by modifying measurements rather than initial poses. It identifies two sufficient consistency conditions—pairwise consistency of relative rotations and translations on each undirected edge, and minimal consistency of the summed offsets—and enforces both by one distributed exchange per edge before optimization (Cristofalo et al., 2020). This is weaker than full cycle consistency. In the context of distributed dynamics, CORD further uses state and velocity to predict neighbors under communication delay, and an ablation study reports divergence when velocity-based neighbor prediction is removed (Shin et al., 11 May 2026).

Robustness to outliers and corrupted measurements appears in several forms. MM-PGO explicitly supports robust loss kernels, including Huber and Welsch, and reports robustness when outliers are pre-filtered by PCM (Fan et al., 2021). The MARL-based framework uses adaptive edge-gating to denoise noisy edges and reports robustness under loop-closure outlier injection (Ghanta et al., 26 Oct 2025). One 2019 abstract describes an “Active Rendezvous” framework that exploits Wi-Fi Channel State Information for on-demand information exchange and outlier rejection in multi-robot PGO, claiming a $R_i\in SO(d)$1 reduction in ground truth pose error from Active Rendezvous and a $R_i\in SO(d)$2 reduction from CSI-aided outlier rejection; however, the accompanying details for the supplied document state that the manuscript text does not actually contain those robotics components (Wang et al., 2019). This discrepancy is itself noteworthy because it cautions against conflating abstract-level claims with the verified technical contents of a manuscript.

Optimality guarantees vary sharply across methods. Many distributed solvers guarantee convergence only to first-order critical points or stationary points under smoothness, Lipschitz-gradient, or bounded-delay assumptions (Tian et al., 2020, Fan et al., 2020, Fan et al., 2021, Chen et al., 2024, Zhao et al., 23 Jun 2026). By contrast, the distributed certifiably correct framework proves exact recovery of the original PGO from a sparse SDP relaxation under moderate noise and adds distributed verification plus saddle escape to certify or improve a candidate solution (Tian et al., 2019). Li et al. also state that the critical point obtained by IRBCD is globally optimal after their multi-level partitioning and low-rank Riemannian treatment (Li et al., 2024). A plausible implication is that “distributed PGO” now spans both heuristic local solvers and globally certifiable relaxation-based pipelines, and the distinction is essential when comparing methods.

5. Performance, scalability, and communication–computation trade-offs

The empirical literature consistently evaluates distributed PGO by objective value, RMSE or APE, iteration count, runtime, and communication burden, but the reported trade-offs differ by algorithmic family.

DGS-based distributed mapping reports convergence to within $R_i\in SO(d)$3 of the centralized two-stage cost in $R_i\in SO(d)$4 iterations in simulations up to 49 robots, communication linear in the number of separators, robustness up to $R_i\in SO(d)$5 noise, and real-world field performance within centimeters and tenths of a degree of centralized Gauss–Newton while using $R_i\in SO(d)$6 KB per rendezvous instead of tens of MB of raw LiDAR frames (Choudhary et al., 2017). These results established a baseline narrative: modest local communication can closely approximate centralized back-end accuracy.

GeoD emphasizes scalability with graph size. It reports that, on average, GeoD converges 20 times more quickly than DGS to a value with 3.4 times less error relative to the global minimum provided by SE-Sync, and on graphs larger than 1000 poses it converges over 100 times faster than DGS (Cristofalo et al., 2020). ASAPP addresses a different bottleneck: synchronization. Under bounded delay, it proves global first-order convergence with a stepsize depending on worst-case delay and graph sparsity, and empirically shows resilience to delays from $R_i\in SO(d)$7 to $R_i\in SO(d)$8 s without divergence, while often matching or improving DGS on benchmark datasets (Tian et al., 2020).

Partition quality and load balancing strongly affect performance. BDPGO reports that repartitioning reduces average imbalance from $R_i\in SO(d)$9–$t_i\in\mathbb R^d$0 to approximately $t_i\in\mathbb R^d$1, decreases edge-cut and communication-volume factors, yields DGS wall-clock speedups of $t_i\in\mathbb R^d$2–$t_i\in\mathbb R^d$3, and maintains coherence under ad-hoc sub-swarm splitting and reunion (Xu et al., 2021). In the multi-level partitioning framework, Highest + IRBCD reduces average iteration count by $t_i\in\mathbb R^d$4–$t_i\in\mathbb R^d$5, cuts per-iteration communication by $t_i\in\mathbb R^d$6–$t_i\in\mathbb R^d$7, and lowers or matches final cost by approximately $t_i\in\mathbb R^d$8–$t_i\in\mathbb R^d$9 relative to DGS and sequential partitioning plus RBCD; on the Garage dataset, the per-robot average runtime drops from $(i,j)$0 ms for DGS and $(i,j)$1 ms for Seq+RBCD to $(i,j)$2 ms for Highest+IRBCD (Li et al., 2024).

ROBO makes the communication–iteration trade-off explicit. Across 22 benchmarks, overlap $(i,j)$3 yields an average inter-robot data cost of approximately $(i,j)$4 Kb per iteration and a worst case of approximately $(i,j)$5 Kb per pair, while reducing the number of iterations by $(i,j)$6 relative to the $(i,j)$7 RBCD baseline. The paper further notes that at $(i,j)$8 Hz, $(i,j)$9 Kb corresponds to approximately $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$0 Mb/s, which it characterizes as well within modern radios rated at $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$1 Mb/s (Sonawalla et al., 3 Mar 2026). This is one of the clearest quantitative statements in the literature that extra communication can be intentionally purchased to reduce iteration count.

More recent methods argue that approximate curvature and problem-specific dynamics can reduce rounds without increasing total burden. DRAN states that approximate-Newton updates reduce iteration count and communication overhead without sacrificing estimation accuracy, and reports near-centralized objective values within $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$2 error while reducing total transmitted data by $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$3–$$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$4 compared with DC2-PGO and DGS (Zhao et al., 23 Jun 2026). CORD reports superior performance against distributed Jacobi-style descent, AMM-PGO, and MESA in both synchronous and asynchronous regimes, with neighbor prediction crucial under packet delay and loss (Shin et al., 11 May 2026).

Learning-based results are reported on a different axis. “Policies over Poses” claims that learned MARL-based actors reduce the global objective by an average of $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$5 more than the state-of-the-art distributed PGO framework while enhancing inference efficiency by at least $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$6, and that actor replication allows scaling to much larger teams without retraining (Ghanta et al., 26 Oct 2025). Because the method is restricted to planar PGO and depends on learned policies, a plausible implication is that these gains are solver-class-specific rather than directly interchangeable with certifiable or geodesic-manifold baselines.

6. Applications, extensions, and research directions

Distributed PGO is primarily motivated by collaborative SLAM, but the surveyed work shows a broader application envelope. Choudhary et al. explicitly connect it to distributed mapping under privacy and communication constraints and extend the framework to object-based map models, where inter-robot factors arise from shared object observations rather than raw scene geometry (Choudhary et al., 2017). DRAN develops this direction into a fully decentralized object-based multi-robot PGO framework on $$\widetilde T_{ij}=(\widetilde R_{ij},\widetilde t_{ij})$$7, separating private trajectories from public object variables and coupling the latter through consensus constraints and biased-prior augmented Lagrangian terms (Zhao et al., 23 Jun 2026). This marks a shift from robot-only trajectory estimation to mixed trajectory–landmark state estimation within the same distributed back end.

Collaborative localization without full SLAM also appears. DGORL uses RSSI-derived ranges, motion models, and distributed graph optimization for multi-robot relative localization over dynamic networks, rather than full feature-based collaborative mapping (Latif et al., 2022). GeoD includes a multi-UAV vision-based SLAM experiment in which seven quadrotors exchange relative poses extracted from onboard images and estimate their trajectories in a fully distributed manner (Cristofalo et al., 2020). These examples show that the distributed PGO abstraction is not limited to one sensing modality.

Several lines of current work expand the design space rather than merely tuning existing solvers. Overlapping decomposition replaces the binary centralized/distributed divide by a tunable continuum (Sonawalla et al., 3 Mar 2026). Continuous Riemannian dynamics reinterpret optimization as damped physical evolution on Lie groups and use velocity states for principled latency compensation (Shin et al., 11 May 2026). Riemannian approximate-Newton methods seek stronger local conditioning without exchanging full Hessians (Zhao et al., 23 Jun 2026). Learned MARL policies replace hand-designed iterative updates by graph-conditioned control policies (Ghanta et al., 26 Oct 2025).

At the same time, the literature identifies recurring limitations. DGORL notes NP-hardness of the global combinatorial choice, sensitivity to poor initial guesses, and the need for extension to three-dimensional motion and real-world Wi-Fi experiments (Latif et al., 2022). ROBO states that increasing overlap raises both communication and local computation, even when convergence accelerates (Sonawalla et al., 3 Mar 2026). ASAPP’s theoretical stepsize becomes smaller as worst-case delay and problem sparsity worsen (Tian et al., 2020). Many methods remain benchmark-driven and rely on partition strategies, initialization quality, or moderate-noise assumptions.

Taken together, the field presents distributed PGO as a mature but still rapidly diversifying area of SLAM back-end research. The shared core is stable—a sparse, manifold-constrained, measurement-graph estimation problem—but the dominant research questions have shifted toward how best to distribute curvature information, certify optimality, survive asynchronous and time-varying communication, balance subproblems, incorporate objects and semantics, and exploit learned priors without losing the structural advantages of graph-based estimation (Tian et al., 2019, Li et al., 2024, Sonawalla et al., 3 Mar 2026, Zhao et al., 23 Jun 2026).