Policies over Poses: Reinforcement Learning based Distributed Pose-Graph Optimization for Multi-Robot SLAM

TL;DR

This work tackles robust, scalable distributed pose-graph optimization for multi-robot SLAM by casting PGO as a cooperative, partially observable Markov game where each robot selects an edge and applies a small pose correction. It couples an edge-conditioned graph neural encoder with an adaptive edge-gate denoising module and a memory-augmented MARL policy (GA-SAC), followed by a consensus step via information-weighted ADMM to reconcile inter-robot constraints. The approach yields substantial improvements over state-of-the-art distributed PGO baselines in both final objective $F(x)$ and inference speed, with additional gains when used as a warm-start for classical solvers; actor replication further enables scalable performance to large teams without retraining. These results demonstrate a practical, robust, and scalable solution for real-time multi-robot SLAM and hold promise for integration with structure-from-motion pipelines and downstream bundle adjustment.

Abstract

We consider the distributed pose-graph optimization (PGO) problem, which is fundamental in accurate trajectory estimation in multi-robot simultaneous localization and mapping (SLAM). Conventional iterative approaches linearize a highly non-convex optimization objective, requiring repeated solving of normal equations, which often converge to local minima and thus produce suboptimal estimates. We propose a scalable, outlier-robust distributed planar PGO framework using Multi-Agent Reinforcement Learning (MARL). We cast distributed PGO as a partially observable Markov game defined on local pose-graphs, where each action refines a single edge's pose estimate. A graph partitioner decomposes the global pose graph, and each robot runs a recurrent edge-conditioned Graph Neural Network (GNN) encoder with adaptive edge-gating to denoise noisy edges. Robots sequentially refine poses through a hybrid policy that utilizes prior action memory and graph embeddings. After local graph correction, a consensus scheme reconciles inter-robot disagreements to produce a globally consistent estimate. Our extensive evaluations on a comprehensive suite of synthetic and real-world datasets demonstrate that our learned MARL-based actors reduce the global objective by an average of 37.5% more than the state-of-the-art distributed PGO framework, while enhancing inference efficiency by at least 6X. We also demonstrate that actor replication allows a single learned policy to scale effortlessly to substantially larger robot teams without any retraining. Code is publicly available at https://github.com/herolab-uga/policies-over-poses.

Figures (5)

• Figure 1: Illustration of the proposed distributed framework. Each robot processes its partitioned subgraph with a GNN encoder equipped with a denoising mechanism, iteratively updates edge poses, and a final consensus step reconciles overlaps into a globally consistent pose graph.
• Figure 2: Overview of the proposed methodology. The -orange represents edge-conditioned GNN encoder, -green denotes adaptive edge-gate denoising, and -violet indicates the action heads in MARL actor.
• Figure 3: Cumulative Learning Efficiency of the Proposed Solution Across 3 Training Environments for n=3 (Left) and n=7 (Right) Actors.
• Figure 4: (a) Initial Noisy Estimate (b) MM-PGO (c) Prop-V1 (d) Prop-V2 results on datasets, where each color represents the trajectory of each robot.
• Figure 5: (Left) Global objective $F(x)$(Right) Runtime versus team size $n$.