riMESA: Consensus ADMM for Real-World Collaborative SLAM
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "riMESA: Consensus ADMM for Real-World Collaborative SLAM".
Dev: Here's the gist of the paper. Collaborative Simultaneous Localization and Mapping (C-SLAM) is a fundamental capability for multi-robot teams as it enables downstream tasks like planning and navigation However, existing C-SLAM back-end algorithms struggle to address the practical realities of real-world deployments (i.e.
Rosa: communication limitations, outlier measurements, and online operation) In this paper we propose Robust Incremental Manifold Edge-based Separable ADMM (riMESA) – a robust, incremental, and distributed CSLAM back-end that is resilient to outliers, reliable in the face of limited communication, and can compute accurate state estimates for a multi-robot team in real-time Through the development of riMESA, we make an argument for the use of Consensus Alternating Direction Method of Multipliers as a theoretical foundation for distributed optimization tasks in robotics like C-SLAM due to its flexibility, accuracy, and fast convergence.
Dev: Problem Formulation and Challenges The generic C-SLAM problem involves a team of robots estimating state variables using noisy measurements A key challenge is the need to solve the robust incremental C-SLAM problem which requires solving the optimization problem at every timestep with access to the previous solution.
Rosa: This formulation is defined as: 1.
Title and authors: Rosa: So we’re looking at this paper today titled riMESA: Consensus ADMM for Real-World Collaborative SLAM, and the authors are Daniel McGann and Michael Kaess. It sounds a bit technical, but basically, they’re tackling the challenge of making collaborative mapping work when things get messy in the real world.
Dev: Yeah, it's about C-SLAM back-ends that struggle with real deployment stuff like bad communication or bad measurements. So they propose this new approach using Consensus Alternating Direction Method of Multipliers as a theoretical base for distributed optimization tasks in robotics, which is pretty flexible and fast to converge.
Taro: What I find interesting right away is the focus on making it robust and incremental, meaning it can handle things step-by-step without needing the whole picture solved all at once.
Rosa: Exactly. They’re proposing a system that can give you accurate state estimates for a multi-robot team in real time, even when communication is limited or measurements are potentially incorrect. It's designed to be resilient under those kinds of conditions.
Dev: The big implication here is moving away from batch solutions where you solve the whole optimization problem once, toward something that updates incrementally and consistently as new information comes in.
The paper's summary: Rosa: So, what exactly is riMESA? It’s described as a robust incremental and distributed CSLAM back-end that handles outliers well by reformulating the problem through M-Estimation. They use an incremental edge-based CADMM process to let robots work incrementally and asynchronously.
Dev: That means when new measurements come in, they run this robust local solver called riSAM to compute updates in real time, which helps manage those outlier measurements effectively. And they also have a communication handler that manages the asynchronous pairwise communications between robots, dealing with latency and failures.
Taro: I see how that handles the communication issue—by having a way to manage what happens when connections drop out or if there's a delay between robots talking to each other. That’s crucial for any field deployment scenario we deal with.
Rosa: Right, so it’s not just solving the problem once; it’s designed to solve the robust incremental C-SLAM problem which requires solving the optimization problem at every time step with access to the previous solution.
Dev: The math they use involves a SIG Kernel update sequence of zero point zero five nine point nine point five one, and they introduce dual variable decay with a decay rate of zero point nine to make sure recent information dominates the dual estimates.
Taro: That sounds like a solid way to manage the trade-off between old and new data influencing the solution. What I wonder about that communication handler—how does it actually manage those failures you mentioned earlier?
Rosa: It involves two stages of communication, first updating each other on shared variables, and then sharing local estimates for the jointly known set of shared variables. They’ve shown that on COSMO-Bench datasets, riMESA is most frequently the best performing method.
The paper's improvements: Dev: The authors put a lot of effort into making it robust by using M-Estimation to downweight the effects of outliers in the state variable update within riSAM. They also introduce an incremental version of Graduated Non-Convexity which efficiently updates only the relevant subproblem at each time step.
Taro: So, when things go wrong with a measurement, it doesn't just blow up the whole calculation; it’s downweighted by this M-Estimation approach. That’s a practical improvement for real-world noise.
Rosa: And they also introduce dual variable decay to downweight old dual estimates so that recent information has more influence on the solution, which is another key piece of their incrementalization strategy.
Dev: The communication handling is also designed to be flexible, allowing asynchronous pairwise communications and managing latency and failures through a Communication Handler that maintains a cached copy of the algorithm’s state, letting robots continue local updates while waiting for inter-robot information.
Taro: That sounds like it builds in some graceful recovery mechanisms rather than just failing when the network hiccups. How does this compare to what we see in other distributed optimization literature?
Rosa: The paper makes a strong argument for using the Consensus Alternating Direction Method of Multipliers as a theoretical foundation because it offers flexibility, accuracy, and fast convergence for these kinds of distributed optimization tasks.
Conclusion: Dev: So to wrap up on riMESA: it’s a robust, incremental, and distributed CSLAM back-end that can compute accurate state estimates in real time under limited communication conditions. It seems they've managed to bridge the gap between theoretical distributed optimization and practical multi-robot deployment needs.
Taro: I think what stands out is how it scales; they show it can handle both large teams and long term operation, which is important because prior works often saw performance degrade at larger scales, so consistency across different sizes is a big deal.
Rosa: Ultimately, riMESA seems to be the only method that performs consistently across all the datasets they tested. It gives you high-quality results whether you’re dealing with a small team or a big one operating over a long time.
Dev: It’s moving us toward solutions for C-SLAM that are not just mathematically sound in isolation, but ones that actually work reliably when deployed on ad hoc, sparse, and unreliable communication networks. That’s the main contribution of riMESA.
Taro: Yeah, it gives us a concrete framework to think about how to make these systems more resilient when the world misbehaves unexpectedly during deployment.
Rosa: We’ve talked about riMESA: Consensus ADMM for Real-World Collaborative SLAM today and that’s all we have for this paper.
Carnegie Mellon University
cs.RO
Submitted: 2026-03-01
Updated: 2026-11-11
Comments: This article has been accepted for publication in IEEE Transactions on Robotics. DOI: 10.1109/TRO.2026.3706569
Code: https://github.com/rpl-cmu/risam-v2
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 90/100
The gist: Collaborative Simultaneous Localization and Mapping (C-SLAM) is a fundamental capability for multi-robot teams as it enables downstream tasks like planning and navigation However…
Key concepts
- Robust Incremental C-SLAM
- This is a method for estimating the positions of multiple robots simultaneously while continuously adding new data and handling errors. It solves the problem step-by-step using previous solutions to be faster and more reliable, specifically designed to cope with noisy measurements and outliers.
- M-Estimation
- M-Estimation is a technique used to make optimization problems resistant to outliers. Instead of treating all data equally, it assigns different weights based on how likely a measurement is. This helps the system ignore or downweight measurements that are clearly incorrect, improving overall accuracy.
- Consensus Alternating Direction Method of Multipliers (CADMM)
- CADMM is a mathematical framework used for distributed optimization where multiple robots work together to agree on a shared state. Robots locally constrain their estimates using 'biased priors' and then iteratively refine these constraints until the entire team converges on a consistent solution.
- Communication Handler
- This component manages the asynchronous and unreliable communication between robots. It allows robots to continue performing local calculations even when communication is delayed or fails, caching necessary information to ensure updates happen correctly once communication is restored.
Terminology
Summary
Collaborative Simultaneous Localization and Mapping (C-SLAM) is a fundamental capability for multi-robot teams as it enables downstream tasks like planning and navigation However, existing C-SLAM back-end algorithms struggle to address the practical realities of real-world deployments (i. p. 0e. communication limitations, outlier measurements, and online operation) In this paper we propose Robust Incremental Manifold Edge-based Separable ADMM (riMESA) – a robust, incremental, and distributed CSLAM back-end that is resilient to outliers, reliable in the face of limited communication, and can compute accurate state estimates for a multi-robot team in real-time Through the development of riMESA, we make an argument for the use of Consensus Alternating Direction Method of Multipliers as a theoretical foundation for distributed optimization tasks in robotics like C-SLAM due to its flexibility, accuracy, and fast convergence. p. 0
Problem Formulation and Challenges
The generic C-SLAM problem involves a team of robots estimating state variables using noisy measurements A key challenge is the need to solve the robust incremental C-SLAM problem which requires solving the optimization problem at every timestep with access to the previous solution. This formulation is defined as:
- Batch C-SLAM seeks to solve for state variables via Maximum A-Posteriori (MAP) inference 2. p. 1 Robust Incremental C-SLAM defines the problem as:
Θ t MAP = arg min Θ t∈Ω t ∑ i∈ ∑ m∈ t i ‖‖ h(Θ t i) − m‖‖ 2 Σm given: Θ t−1 MAP.
The problem is further complicated by the Maximum-Consensus problem, which is NPHard and generally intractable. p. 1
Algorithm Design and Framework
The paper proposes riMESA as a robust, incremental, and distributed CSLAM back-end that is resilient to outliers. riMESA incrementallyizes an edge-based CADMM process to allow robots to operate incrementally and asynchronously. p. 6 It leverages a robust incremental local solver (riSAM) to address outlier measurements and compute updates in real-time. p. 0 The algorithm consists of running Alg. p. 6 3 when new measurements are added, Alg. p. 8 4 when communications are initialized, and Alg. p. 8 5 when communications are successfully completed. p. 9
Robustness and Incrementalization Techniques
The robustness of riMESA is achieved by reformulating the problem through the lens of M-Estimation to downweight the effects of outliers. The state variable update in riSAM uses a SIG Kernel with a control parameter update sequence μ = [0. p. 90, 0.5, 0.9, 0.95, 1.0] to ensure that factors classified as outliers are unable to negatively affect the solution. p. 9 Furthermore, dual variable decay is introduced with a decay rate δ = 0. p. 79 to downweight old dual estimates so that recent information dominates the dual’s influence. p. 8
Communication Handling and Real-World Performance
riMESA incorporates a Communication Handler
to manage communication between robots, allowing for asynchronous pairwise communications and handling latency and failures. The two-stage communication process involves first updating each other on shared variables and then sharing local estimates for the jointly known set of shared variables. p. 9 In real-world experiments on COSMO-Bench datasets, riMESA is most frequently the best-performing method. p. 0
Experimental Validation and Conclusion
The performance evaluation on synthetic and real-world data demonstrates that riMESA is able to generalize across conditions, produce accurate state estimates, operate in real-time, and outperform the accuracy of prior works by a factor >7x on real-world datasets. The results show that riMESA can scale to both large teams and long-term operation, providing consistent, high-quality results across scales. p. 2 Ultimately, riMESA is the only method that performs consistently across all the datasets. p. 0
--- Page 1 ---
riMESA: Consensus ADMM for Real-World Collaborative SLAM Daniel McGann and Michael Kaess riMESA estimates the state of a multi-robot team (,,) from noisy, potentially incorrect measurements (•,•,•) using only sparse, unreliable communication (O) riMESA is a C-ADMM-based distributed optimization algorithm in which robots locally constrain shared state using “biased priors” (•) Over time, as communication is available, riMESA tightens equality constraints with dual variables (lambda) to provide consistent solutions for the team Meanwhile, robots incorporate new measurements efficiently using the riSAM algorithm, which handles potential outlier measurements (◑) using M-Estimations and an incremental version of Graduated Non-Convexity, which efficiently updates only the relevant subproblem (◼) at each timestep
--- Page 2 ---
2
We seek to solve the generic C-SLAM problem in which a team of robots estimates state variables Θ using noisy measurements . Importantly, some measurements taken by robots will be inter-robot measurements Due to inter-robot measurements, multiple robots may observe the same variable There are two key challenges when utilizing the C-SLAM problem formulation for real-world deployments. p. 1 Firstly, during real deployments, we are not interested in just solving this optimization problem once. p. 1 p. 1 Secondly, this NLS formulation is inherently sensitive to outliers. p. 1 p. 1
--- Page 3 ---
A. Batch C-SLAM The de-facto standard for SLAM is to formulate the problem as a factor-graph and solve for the state variables via Maximum A-Posteriori (MAP) inference B. p. 1 Robust Incremental C-SLAM There are two key challenges when utilizing the C-SLAM problem formulation for real-world deployments. p. 1 Firstly, during real deployments, we are not interested in just solving this optimization problem once. Secondly, this NLS formulation is inherently sensitive to outliers. Combining these challenges, we define the robust incremental C-SLAM problem:
Θ t MAP = arg min Θ t∈Ω t ∑ i∈ ∑ m∈ t i ‖‖ h(Θ t i) − m‖‖ 2 Σm given: Θ t−1 MAP.
--- Page 4 ---
C. Maximum-Consensus An important nuance in the definition of (3) is that the set of true
inliers is unknowable. A knowable and, therefore, more practical definition we can use to define inliers is that of “jointconsistency. p. 1” We define a set of measurements as jointly consistent if there exists some state of the world Θ such that:
‖‖ h(Θ) − m‖Σm < chi2 m(T) ∀ m ∈ (5).
--- Page 5 ---
D. C-SLAM Communication In the C-SLAM setting, information is distributed across the team, and robots must communicate to solve the problems discussed above. p. 1 In real-world applications, we can only reliably assume that robots can communicate over an ad-hoc network built from the robots themselves. p. 1
--- Page 6 ---
E. Problem Scope To support real-world deployments of multi-robot teams, we require algorithms that can solve the robust incremental C-SLAM problem (3) even when robots have access to only ad-hoc, sparse, and unreliable communication (Sec. p. 2 II-D).
--- Page 7 ---
III. RELATED WORK A. p. 2 C-SLAM Optimization Prior works have not provided an effective solution to the incremental, robust C-SLAM problem with adhoc, sparse, and unreliable communication. p. 2 Some prior works have proposed algorithms based on Multi-Block ADMM (MB-ADMM [39]), Distributed GaussSeidel (DGS [7]), Distributed Gradient-Descent (GeoD [8]), Majorization-Minimization (MM-PGO [11, 12]), and Distributed Riemannian Gradient Descent (DC2-PGO [10], ASAPP [9]). p. 2
B. Robust Optimization Prior works have proposed methods to address outlier measurements by approximately solving the Maximum-Consensus problem
The paper discusses approaches such as M-Estimation, Relaxed Consensus, Variable Augmentation, and Continuation. Some prior works have proposed using M-Estimation in their optimization. p. 2
C. Robust C-SLAM Optimization Prior works have looked to make C-SLAM optimization robust by combining existing back-ends with methods for robust optimization
Numerous centralized and decentralized methods have proposed incorporating existing robust optimization methods. Distributed systems have taken similar approaches. p. 3
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Robust incremental C-SLAM back-end riMESA can compute accurate state estimates for a multi-robot team in real-time, operating in real-world deployments with limited communication by incrementally solving an edge based consensus ADMM process.
-
The system can handle outlier measurements by leveraging MEstimation, which is integrated into the local optimization step via riSAM to address the sensitivity of NLS formulations to outliers.
-
The algorithm supports asynchronous communication and handles latency and communication failures through a Communication Handler that maintains a cached copy of the algorithm’s state, allowing robots to continue local updates while waiting for inter-robot information.
-
A mechanism called Robust Weighted Biased Priors (RWBP) allows robots to effectively consider the constraints that biased priors represent as
out-of-date,
enabling them to better utilize new information to correct misclassifications in incremental settings. -
The system can be scaled to both large teams and long-term operation, providing consistent, high quality results across different scales, unlike prior works which see degraded performance at large scales.
Sources
- GeoD: Consensus-based Geodesic Distributed Pose Graph Optimization
- SlideSLAM: Sparse, Lightweight, Decentralized Metric-Semantic SLAM for Multi-Robot Navigation
- A micro Lie theory for state estimation in robotics
- A Riemannian ADMM
- COSMO-Bench: A Benchmark for Collaborative SLAM Optimization
- CU-Multi: A Dataset for Multi-Robot Collaborative Perception
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation