riMESA: Robust Incremental Manifold SLAM

Updated 5 July 2026

riMESA is a distributed C-SLAM back-end that performs robust, incremental pose graph optimization on SE(2)/SE(3) manifolds, addressing challenges like limited communication and outlier measurements.
It integrates consensus ADMM with edge-based separability so that only shared variables between robots are updated asynchronously to reduce message sizes and ensure real-time performance.
By leveraging robust techniques such as riSAM with continuation and dual decay, riMESA achieves superior accuracy—over 7× improvement—on real-world datasets compared to prior methods.

riMESA, short for Robust Incremental Manifold Edge-based Separable ADMM, is a distributed back-end for collaborative simultaneous localization and mapping (C-SLAM) designed for multi-robot teams operating under real deployment constraints such as limited communication, outlier measurements, and online incremental operation. It formulates multi-robot pose-graph optimization on manifolds such as $SE(2)$ and $SE(3)$ , combines consensus ADMM with robust incremental local optimization, and enforces agreement only on shared variables through edge-based separable updates. The method is presented as a robust, incremental, and distributed C-SLAM back-end that is resilient to outliers, reliable under limited communication, capable of real-time state estimation, and empirically able to outperform the accuracy of prior distributed methods by a factor $>7\times$ on real-world datasets (McGann et al., 1 Mar 2026).

1. Definition and problem setting

riMESA addresses the multi-robot C-SLAM back-end problem. Each robot $k \in \{1,\ldots,K\}$ estimates its own trajectory and possibly a shared map, while inter-robot measurements couple the local estimation problems. The global state consists of robot poses $x=\{x_i\}$ , with $x_i \in SE(3)$ in the general formulation, and edges in the pose graph encode relative transforms from odometry and loop closures, including both intra-robot and inter-robot factors (McGann et al., 1 Mar 2026).

The method is explicitly motivated by three classes of real-world constraints. First, communication is assumed to be ad-hoc, sparse, unreliable, asynchronous, and neighbor-to-neighbor only, with latency, frequent disconnections, and “Two Generals” failures. Second, loop-closure measurements may contain outliers induced by perceptual aliasing, so the estimator must be robust without assuming batch global verification. Third, the system is intended for online operation, where new measurements arrive over time and the back-end must maintain low latency for downstream planning and navigation (McGann et al., 1 Mar 2026).

riMESA’s stated objectives are therefore fourfold: robustness through M-estimation and continuation, incremental real-time behavior via amortized consensus and warm starts, distributed consensus on shared variables through edge-based separable ADMM, and accuracy with fast convergence under realistic communication constraints. This combination places riMESA within distributed nonlinear least-squares and factor-graph optimization, but with an emphasis on asynchronous field operation rather than idealized synchronized consensus (McGann et al., 1 Mar 2026).

A common misunderstanding is to treat riMESA as a full C-SLAM stack. In the cited formulation it is specifically a back-end; front-end tasks such as distributed loop-closure generation, timestamp-based direct measurement matching, and descriptor-based landmark matching remain distinct components of the overall system (McGann et al., 1 Mar 2026).

2. Manifold formulation of the collaborative pose graph

The underlying estimation problem is a robust pose-graph optimization on $SE(3)$ or $SE(2)$ . For an edge $(i,j)$ , riMESA uses a manifold residual defined through the logarithm map: $r_{ij}(x_i,x_j)=\operatorname{Log}\!\big(T_{ij}^{-1}\,\big(x_i^{-1}x_j\big)\big)\in\mathbb{R}^6.$ Here $SE(3)$ 0 is the measured relative transform and $SE(3)$ 1 is its covariance (McGann et al., 1 Mar 2026).

The robust factor-graph objective is

$SE(3)$ 2

where $SE(3)$ 3 is an M-estimator. The cited formulation lists Huber, Cauchy, Geman–McClure, and a continuation kernel as examples (McGann et al., 1 Mar 2026).

The paper also gives an optional switchable-constraint formulation,

$SE(3)$ 4

but explicitly states that riMESA does not use explicit switches. Instead, it relies on robust M-estimation, continuation via riSAM, and robust handling of consensus priors (McGann et al., 1 Mar 2026).

This choice is significant because it shifts robustness away from discrete edge-selection variables toward a continuously reweighted incremental optimization procedure. A plausible implication is that the design prioritizes stable online updates and reduced combinatorial overhead over explicit binary loop-closure gating, which is consistent with the observation that batch maximum-consensus is NP-hard (McGann et al., 1 Mar 2026).

3. Consensus ADMM and edge-based separability

riMESA uses consensus ADMM as the distributed optimization scaffold. The global problem is partitioned by robot: robot $SE(3)$ 5 maintains local variables $SE(3)$ 6 and local factors $SE(3)$ 7, while inter-robot couplings are represented through shared states and consensus variables. In the constrained consensus form,

$SE(3)$ 8

where $SE(3)$ 9 denotes consensus variables for states shared between robot $>7\times$ 0 and its neighbors (McGann et al., 1 Mar 2026).

The augmented Lagrangian is

$>7\times$ 1

The canonical ADMM updates are then written as a local $>7\times$ 2-update, a consensus $>7\times$ 3-update, and a dual update: $>7\times$ 4

$>7\times$ 5

$>7\times$ 6

with $>7\times$ 7 as the scaled dual (McGann et al., 1 Mar 2026).

riMESA adapts this framework to manifolds and to sparse shared-state structure. The central architectural decision is edge-based separability: consensus constraints are imposed only on shared variables, so only robots connected by an edge $>7\times$ 8 participate in the corresponding update. Equality on $>7\times$ 9 is enforced through geodesic constraints based on logarithmic errors, and the per-edge consensus variable is updated in closed form using geodesic SPLIT interpolation at $k \in \{1,\ldots,K\}$ 0: $k \in \{1,\ldots,K\}$ 1 This yields small message sizes and permits asynchronous pairwise communication rather than synchronized global rounds (McGann et al., 1 Mar 2026).

A frequent misconception is that ADMM-based distributed SLAM necessarily requires network-wide barrier synchronization. In riMESA, the update pattern is explicitly asynchronous and pairwise: only the communicating edge is updated, and no global round structure is required (McGann et al., 1 Mar 2026).

4. Robustness, incrementality, and local optimization

riMESA’s robustness is built around riSAM (Robust Incremental SAM). All measurement residuals are wrapped by a robust loss, and riSAM performs continuation or graduated non-convexity incrementally. The kernel control parameter $k \in \{1,\ldots,K\}$ 2 is scheduled so that outlier downweighting becomes less sensitive to initialization than with fixed M-estimators (McGann et al., 1 Mar 2026).

A second robustness mechanism is the use of robust weighted biased priors (RWBP) for the consensus terms themselves. The ADMM quadratic-plus-dual priors are wrapped by robust kernels so that out-of-date consensus constraints can be effectively rejected and later reclassified as the estimate evolves. The paper reports representative prior-noise balancing values of $k \in \{1,\ldots,K\}$ 3 rad for rotation and $k \in \{1,\ldots,K\}$ 4 m for translation (McGann et al., 1 Mar 2026).

To limit the influence of stale information, riMESA applies dual decay: $k \in \{1,\ldots,K\}$ 5 with $k \in \{1,\ldots,K\}$ 6, so that recent information dominates after constraint rejection or abrupt estimate changes (McGann et al., 1 Mar 2026).

Incrementality is equally central. New factors are added over time, and riMESA uses riSAM’s Bayes tree to update only impacted subgraphs, in a manner described as analogous to iSAM2 but robust. The method performs only a single robust local update per measurement, while consensus tightening is amortized over time whenever communication becomes available. Warm starts and trust-region steps based on DogLeg, with Levenberg–Marquardt noted as optional, are used to maintain real-time performance and stability (McGann et al., 1 Mar 2026).

The continuation schedule is given as an example by

$k \in \{1,\ldots,K\}$ 7

and the penalty terms are initialized with $k \in \{1,\ldots,K\}$ 8 and $k \in \{1,\ldots,K\}$ 9, remaining constant thereafter to avoid over-constraining during reclassification (McGann et al., 1 Mar 2026).

These mechanisms distinguish riMESA from distributed formulations that simply attach unmodified quadratic consensus penalties to local optimizers. In riMESA, even the consensus prior is treated as a robust statistical object, reflecting the fact that under lossy asynchronous communication a stale consensus term can behave like an outlier (McGann et al., 1 Mar 2026).

5. Communication model, asynchronous operation, and implementation

The communication model assumes an ad-hoc, bandwidth-limited, lossy, and asynchronous network. Robots communicate pairwise when in range, connections may drop during exchange, and “Two Generals” failures are explicitly part of the design model (McGann et al., 1 Mar 2026).

Communication proceeds in two stages. In Stage 1, robots exchange sets of shared variables, variables needing initialization, local observability flags, and optionally environment descriptor sets for landmark matching. In Stage 2, they exchange current estimates for jointly known shared variables. Typical payloads are described as small, consisting of a few IDs plus 6D $x=\{x_i\}$ 0 minimal coordinates per shared variable, with optional compression or quantization such as float16 or delta coding (McGann et al., 1 Mar 2026).

Under limited communication, riMESA uses several tactics. Asynchronous ADMM updates only the active communication edge. Stale information handling computes $x=\{x_i\}$ 1 and $x=\{x_i\}$ 2 from cached data from the moment a communication event began so that both sides use consistent inputs; changes that occur during latency are absorbed at the next exchange. Partial updates are allowed when only some shared variables are known or matched, and Two-Generals failures are handled by later reconciliation, with RWBP and dual decay mitigating temporary asymmetry (McGann et al., 1 Mar 2026).

The implementation loop is likewise event-driven. On receipt of new factors, the system performs bookkeeping for new shared variables and dual caches, executes a local robust riSAM update, and marks variables for re-elimination or re-convexification if communication affected them. On communication completion, it identifies jointly known shared variables, initializes newly shared states using local and owner estimates according to observability flags, performs the SPLIT-based $x=\{x_i\}$ 3-update, applies the dual update with decay, and marks affected variables for the next local update (McGann et al., 1 Mar 2026).

The manifold representation uses 6D minimal tangent coordinates for $x=\{x_i\}$ 4, with Exp/Log retractions, geodesic pose errors, and spherical interpolation on rotation plus linear interpolation on translation in SPLIT. Sparse Cholesky and Bayes tree machinery support local re-elimination of affected cliques, and only variables moved beyond thresholds are re-linearized (McGann et al., 1 Mar 2026).

The cited implementation guidance also notes that riSAM is open-source in v2, and that COSMO-Bench provides realistic multi-robot LiDAR data and communication models suitable for evaluation (McGann et al., 1 Mar 2026).

6. Convergence, complexity, evaluation, and limitations

The convergence discussion is deliberately qualified. In convex consensus problems, ADMM has linear convergence, and edge-based asynchronous ADMM converges at $x=\{x_i\}$ 5 in convex settings. For the nonconvex manifold problems arising in C-SLAM with nonlinear constraints and asynchrony, the paper states that general theoretical guarantees are limited; riMESA instead reports empirical convergence across a broad range of conditions, aided by robust continuation and trust-region steps (McGann et al., 1 Mar 2026).

Per robot, the dominant cost is the local robust incremental riSAM solve, whose complexity depends on the size of the local factor graph and the affected Bayes-tree subtree. The paper characterizes consensus $x=\{x_i\}$ 6- and $x=\{x_i\}$ 7-updates as $x=\{x_i\}$ 8, with closed-form SPLIT and Log operations that are negligible compared with local solves. Real-time performance is reported as compatible with per-keyframe latency bounds such as approximately $x=\{x_i\}$ 9– $x_i \in SE(3)$ 0 s between keyframes, often with sub-100 ms per update on standard hardware (McGann et al., 1 Mar 2026).

The empirical evaluation spans both synthetic and real-world scenarios. Synthetic experiments include teams of up to 6 robots, 1000-pose trajectories, planar and unconstrained 3D settings, odometry plus intra- and inter-robot loop closures, as well as range-only, bearing-range, and landmark measurements, with 10–25% outliers and multiple noise levels. Real-world evaluation uses COSMO-Bench datasets, described as 24 LiDAR-based multi-robot sequences with realistic communication models, together with Nebula sequences, under communication rates of 5 Hz and latency $x_i \in SE(3)$ 1 ms with realistic Wi-Fi and Pro-Radio success and range maps (McGann et al., 1 Mar 2026).

The reported metrics are ATE after Umeyama alignment, incremental ATE (iATE) weighted by time, and F1 score for inlier/outlier classification when applicable. riMESA is reported to consistently outperform distributed baselines such as DLGBP and DDF-SAM2, and to outperform the ablative kiMESA variant based on iSAM2 with a fixed M-estimator, while approaching centralized oracle robustness without centralized communication (McGann et al., 1 Mar 2026).

On COSMO-Bench, riMESA’s average optimality gap relative to centralized GNC is reported as approximately 45%, whereas DLGBP and DDF-SAM2 are described as having approximately $x_i \in SE(3)$ 2– $x_i \in SE(3)$ 3 worse gaps; this is the basis for the statement that riMESA yields $x_i \in SE(3)$ 4 accuracy improvement on real-world datasets (McGann et al., 1 Mar 2026). The paper further reports that higher communication rates and ranges improve performance in low-SNR settings such as range-only or range-aided problems, whereas in high-SNR C-PGO the method remains robust even under poor communication (McGann et al., 1 Mar 2026).

The principal limitations are also explicit. riMESA remains subject to the general difficulties of nonconvex SLAM; catastrophic front-end misassociations can still trigger temporary divergence, even though robust continuation mitigates this risk. Extremely low SNR and very sparse communication can slow convergence. The authors also identify the lack of comprehensive theory for nonconvex manifold consensus under asynchrony as an open problem (McGann et al., 1 Mar 2026).

7. Nomenclature and disambiguation

The acronym riMESA in the cited robotics paper denotes Robust Incremental Manifold Edge-based Separable ADMM (McGann et al., 1 Mar 2026). The provided source set also contains two unrelated, similarly named methods from wireless communications, which can create bibliographic ambiguity.

Acronym	Expansion	Domain
riMESA	Robust Incremental Manifold Edge-based Separable ADMM	Collaborative SLAM (McGann et al., 1 Mar 2026)
RISMA	Reconfigurable Intelligent Surfaces Enabling Beamforming for IoT Massive Access	RIS-aided mmWave massive IoT (Mursia et al., 2020)
RIMSA	Reconfigurable Intelligent Metasurface Antennas	Multi-user downlink metasurface arrays (Wei et al., 23 Jun 2025)

RISMA is a beyond-5G downlink framework that jointly optimizes base-station precoding and RIS parameters to minimize system sum mean squared error in massive IoT access (Mursia et al., 2020). RIMSA, by contrast, refers to a radiative metasurface antenna architecture for multi-user downlink with joint digital and metasurface optimization (Wei et al., 23 Jun 2025). Neither is related to collaborative SLAM.

This nomenclature overlap has a practical consequence for literature search and citation hygiene: the robotics method should be identified by the full expansion or by its arXiv identifier when precision is required. A plausible implication is that disambiguation is especially important in cross-domain indexing systems, because riMESA, RISMA, and RIMSA occupy distinct technical lineages despite near-homographic acronyms (McGann et al., 1 Mar 2026).