Learning Adaptive Solvers for Distributed Factor Graph Optimization on Matrix Lie Groups

Noosaga Placements

• Robot Perception And Slamsubfield95%
  The paper directly addresses distributed factor graph optimization for SLAM, specifically pose graph optimization (PGO) and projective submap alignment, which are core SLAM problems. The method is evaluated on SLAM benchmarks and compared to SLAM-specific baselines.
  Modern robotic perception increasingly involves large-scale geometric optimization problems distributed across multiple robots or sessions.Extensive experiments on real-world SE(3) pose graph optimization and SL(4) projective submap alignment show that our method achieves lower objective values than existing distributed baselines on most benchmarks across realistic operating scenarios.State-of-the-art multi-robot SLAM systems use PGO to correct accumulated odometry drift and improve global consistency of trajectory estimation
• Graph-based SLAMframework90%
  The paper is situated within the Graph-based SLAM framework, as it addresses factor graph optimization for SLAM, which is the core of graph-based SLAM. The method is evaluated on pose graph optimization, a canonical graph-based SLAM problem.
  Factor graphs [7, 11] provide a common abstraction for these problems and support optimization over different matrix Lie groups, including rigid motions for pose graph SLAM [41]In this work, we consider distributed factor graph optimization problems with pairwise relative measurements between group-valued variables. This class of factor graphs forms the backbone for modern SLAM and 3D reconstruction systems [7, 11].Extensive experiments on real-world SE(3) pose graph optimization and SL(4) projective submap alignment show that our method achieves lower objective values than existing distributed baselines on most benchmarks across realistic operating scenarios.
• Learning-based SLAMframework85%
  DeepCORD is a learning-augmented framework that uses a neural network policy to adapt solver parameters, making it a learning-based approach to SLAM optimization. It directly uses learning to improve the optimization process.
  we present DeepCORD, a learning-augmented framework for distributed factor graph optimization on general matrix Lie groups.By unfolding a parallel and accelerated Riemannian optimizer into differentiable iterations, DeepCORD learns a self-supervised feedback policy that dynamically adapts solver parameters according to the optimization phase and communication status.We design a self-supervised approach to train DeepCORD via deep unfolding, by minimizing unrolled objectives without relying on supervision from optimal solutions.
• Multi Robot Systemssubfield80%
  The method is designed for distributed optimization across multiple robots, with explicit handling of communication regimes (synchronous/asynchronous) and data locality. The experiments involve multi-robot scenarios.
  many multi-robot, multi-device, and multi-session systems naturally produce factor graphs that are partitioned across robots or devicesIn the distributed setting, the nodes are partitioned across a set of robots A = {1, . . . , N}.DeepCORD retains the communication efficiency of CORD and each iteration uses a single round of communication.
• Machine Learningsubfield70%
  The paper uses learning-to-optimize (L2O) and deep unfolding, which are machine learning techniques. The method involves training a neural network policy via self-supervised learning.
  Inspired by learning-to-optimize (L2O) frameworks [8], we propose DeepCORD, a learning-augmented framework for distributed factor graph optimization on matrix Lie groups.We design a self-supervised approach to train DeepCORD via deep unfolding, by minimizing unrolled objectives without relying on supervision from optimal solutions.The policy πθ processes the graph attributes using a graph encoder followed by a MLP parameter head.
• Multi-Agent Robot Learningframework60%
  The method involves multiple robots learning to coordinate optimization, which relates to multi-agent robot learning. However, the learning is not multi-agent in the sense of each robot learning a separate policy; rather, a shared policy is used across robots.
  The weights θ are shared across all robots and problem instances, enabling generalization across graph sizes, topologies, and communication regimes.At every iteration k, each robot i applies this shared policy to local information to adjust its own parameters

Abstract

Modern robotic perception increasingly involves large-scale geometric optimization problems distributed across multiple robots or sessions. However, existing distributed solvers often depend on brittle hand tuning and primarily target rigid body pose graphs. To address this, we present DeepCORD, a learning-augmented framework for distributed factor graph optimization on general matrix Lie groups. By unfolding a parallel and accelerated Riemannian optimizer into differentiable iterations, DeepCORD learns a self-supervised feedback policy that dynamically adapts solver parameters according to the optimization phase and communication status. The resulting method enables adaptive distributed optimization over matrix Lie groups under both synchronous and asynchronous communication regimes. Extensive experiments on real-world $\mathrm{SE}$(3) pose graph optimization and $\mathrm{SL}$(4) projective submap alignment show that our method achieves lower objective values than existing distributed baselines on most benchmarks across realistic operating scenarios.

Paper Context