CGS-SLAM: Collaborative Gaussian Splatting based SLAM for Multi-Agent Reconstruction

Jean-Daniel de Ambrogi; Aladine Chetouani; Vincent Nguyen; Aurélien Chateigner

cs.CVAug 27, 2026

CGS-SLAM: Collaborative Gaussian Splatting based SLAM for Multi-Agent Reconstruction

Authors: Jean-Daniel de Ambrogi, Aladine Chetouani, Vincent Nguyen, Aurélien Chateigner

Organizations: Université Sorbonne Paris Nord, L2TI, UR 3043, F-93430, Villetaneuse, France · SAS IMPACT, Orléans, France · Université d’Orléans, INSA CVL, LIFO, UR 4022, Orléans, France

Abstract

Recent advances in SLAM have leveraged 3DGS for photorealistic reconstruction and novel view synthesis. However, most methods rely on RGB-D input, which is unavailable on consumer-grade smartphones, and few integrate 3DGS within a collaborative framework. Therefore, we present CGS-SLAM, a hybrid decentralized/centralized system enabling multi-agent 3DGS SLAM using only RGB and inertial data. Each agent performs local tracking with inertial data as a motion prior and reconstructs a scaled map using a metric monocular depth estimator (Depth Pro). Keyframe encodings are shared among agents, enabling dynamic keyframing in regions of spatial overlaps with other agents, enhancing submap alignment. Afterwards, a central server aligns submaps using VGGT as a view alignment model. This bidirectional communication keeps communication cost low during mapping and global reconstruction in difficult GNSS-denied environments. Experiments on multiple datasets demonstrate competitive tracking performance, improved rendering quality over state-of-the-art methods, and accurate submap alignment.

Figures & tables

Figure 1: The global architecture of our framework. Our method allows multiple agents to evolve into an unknown environment and locally reconstruct it. The keyframes encodings are sent to the server, which broadcast them to all agents. Afterwards all agents share their local maps to the server which merge them into a single global map.

Figure 2: Comparison of the rendering between LiDAR (left), MiDaS [ 34 ] depth estimator (middle) and Depth-Pro (right) based Gaussian maps on scene fast-straight from UT-MM. Our method exhibits reduced linear distortion along the horizontal axis. Even when compared against LiDAR measurements, our method demonstrates superior consistency against sensor artifacts found in the left panel.

Figure 3: Raw trajectory output (TUM, fr1/desk), i.e . when results are not rotated, scaled nor translated to fit at best the ground truth, as usually done to evaluate ATE-RMSE. We only align the origin. Our method offers a trajectory almost at scale with the ground truth, while MM3DGS is way off with an overall size of a third of the GT.

Method	Map	Setup	fr1/desk	fr1/desk2	fr1/room	fr2/xyz	fr3/office	Avg
iMAP [ 43 ]	Neural	RGBD	4.9*	–	–	2.0*	58*	–
NiceSLAM [ 56 ]	Neural	RGBD	4.26*	4.99*	34.49*	31.73*	30*	21.09
CP-SLAM [ 16 ]	Neural	RGBD	7.84*	–	–	3.93*	–	–
Go-SLAM [ 54 ]	Neural	RGB	1.6*	2.8*	5.2*	1.0*	2.23	2.56
GS-SLAM [ 50 ]	3DGS	RGBD	3.3*	–	–	1.3*	–	–
MAGIC-SLAM [ 51 ]	3DGS	RGBD	4.71	3.91	14.45	1.15	15.64	7.97

Table 1: Tracking results against SOTA neural methods (top) and 3DGS based methods (bottom) on TUM Dataset (ATE-RMSE

\downarrow

cm). Results marked with * are provided by the original paper. Our method does not use depth measurements and still performs as well as the RGBD-based 3DGS SLAM methods. The first , second , and third ranks are highlighted accordingly on 3DGS based SLAM; Go-SLAM provides the best results on all scenes.

Figure 4: From left to right: RGB GT, RGB rendering, Depth GT, Depth rendering and Depth predicted from fr1/desk (TUM) and Fast-Straight (UT-MM) scenes with our method on monocular settings. We achieve clear rendering fidelity and spatial consistency with the GT.

	Magic-SLAM			MAC-Ego3D			MM3DGS			Ours
Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Square-1	18.46	0.77	0.37	20.60	0.18	0.76	16.06	0.57	0.44	18.13	0.64	0.43
Square-2	–	–	–	20.84	0.17	0.78	17.50	0.60	0.42	18.08	0.62	0.39
Ego-centric	19.10	0.79	0.30	14.38	0.39	0.61	22.81	0.82	0.22	23.34	0.86	0.20
Fast-Straight	22.45	0.86	0.28	21.04	0.20	0.77	21.58	0.78	0.29	24.51	0.82	0.26
fr1/desk	16.57	0.68	0.47	15.66	0.67	0.38	14.73	0.56	0.60	19.44	0.70	0.38

Table 2: Rendering comparison on TUM and UT-MM datasets. PSNR

\uparrow

, SSIM

\uparrow

and LPIPS

\downarrow

. Magic-SLAM and MAC-Ego3D are RGBD, while MM3DGS and ours are RGB-only. We perform better than MM3DGS on all scenes, and achieve similar quality with the RGBD methods. Magic-SLAM failed on Square-2 scene.

Method	Square1		Ego-centric		Square2		Fast-straight		Avg
	ATE	PSNR	ATE	PSNR	ATE	PSNR	ATE	PSNR	ATE	PSNR
MM3DGS (CV)	59.48	16.0	4.09	23.15	67.20	17.51	25.78	21.71	39.14	19.59
Ours (CV)	55.43	17.0	5.02	23.52	62.84	17.96	12.85	21.78	34.03	20.06
MM3DGS (IMU)	47.05	16.06	3.52	22.81	68.50	17.50	16.78	21.58	33.96	19.49
Ours (IMU)	42.33	18.13	4.26	23.34	55.24	18.08	6.07	24.51	26.98	21.02

Table 3: Contribution of the IMU pre-integration compared to the Constant Velocity (CV) Assumption during the tracking phase. All methods are in monocular settings, our method outperforms MM3DGS in most cases. ATE RMSE is in cm and PSNR is in dB. The first , second , and third ranks are highlighted accordingly.

Figure 5: Submap alignment comparison on different scenes association from the UT-MM dataset. Each trajectory is plotted after transformed by the server sub-map alignment process. Our method achieves sub-meter spatial alignment with the ground truth (translation error < 0.40 m), and can work with multiple ( > 2) agents exploring the same scene.

Scenes	with Dynamic KF	without Dynamic KF
Sq1&EgoCtr	37.33	53.40
Sq1&FastStr	18.12	19.08
FastStr&Sq1&EC	39.63	68.13

Table 4: Dynamic KF impact on ATE (cm) on global map. Enabling it brings improved alignment on all scenes. It has limited contribution on Sq1&FastStr as the second trajectory is rectilinear and forward, thus all the keyframes carry the same information

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 10 — Figure S1 : Our communication protocol uses a custom message architecture enabling efficient, low-bandwidth communication of keyframe encodings.

Mapping	Value
Learning Rates (LR)
$\lambda_{D}$	0.05
$\lambda_{C}$	0.75
$\lambda_{S}$	0.2
Camera $t$ / $q$ LR	$0.001$ / $0.003$
Feature / Opacity LR	$0.0025$ / $0.05$

Appendix

Table S1: System parameters used in our experimental setup.

	Depth Pro				Depth Anything V3
Scene	ATE-RMSE	PSNR	SSIM	LPIPS	ATE-RMSE	PSNR	SSIM	LPIPS
Square-1	42.33	18.13	0.64	0.43	61.42	16.42	0.53	0.45
Square-2	55.24	18.08	0.62	0.39	X	X	X	X
Ego-centric	4.26	23.34	0.86	0.20	26.55	18.62	0.69	0.35
Fast-Straight	6.07	24.51	0.82	0.26	12.63	20.54	0.68	0.48
Slow-Straight-2	5.01	25.85	0.84	0.24	6.18	25.66	0.84	0.24

Appendix

Table S2 : Tracking (ATE-RMSE

\downarrow

) and Rendering (PSNR

\uparrow

, SSIM

\uparrow

and LPIPS

\downarrow

) comparison between Depth Pro (DP) and Depth Anything V3 (DA3) setup on TUM and UT-MM datasets. DA3 is competitive and occasionally better on the small TUM sequences, but degrades sharply on the large UT-MM scenes and fails on Square-2, which is why we retain the dual-model setup

Figure S2 : Trajectory comparison between Depth Pro and Depth Anything configuration. We can see that the Depth Pro configuration achieves better tracking, with better handling of turns.

Scene	Agent	ATE-RMSE	PSNR	SSIM	LPIPS
Apt-0	0	12.33	26.96	0.79	0.29
Apt-0	1	15.06	25.20	0.81	0.24
Apt-1	0	199.39	12.92	0.55	0.49
Apt-1	1	122.20	17.65	0.62	0.43
Apt-2	0	19.54	25.20	0.77	0.28
Apt-2	1	150.09	18.83	0.72	0.34

Appendix

Table S3 : Tracking (ATE-RMSE

\downarrow

cm) and Rendering (PSNR

\uparrow

dB, SSIM

\uparrow

and LPIPS

\downarrow

) on Replica Dataset under the constant velocity hypothesis (no IMU).

Figure S3 : Top view of trajectories from global alignment of submaps from ReplicaMultiagent. The tracking error from the lack of IMU contaminates the submap alignments, resulting in poor ATE-RMSE.

Figure S4 : Rendering examples of the scenes from Apt-0 and Apt-1 of the Replica dataset.

Figure S5 : Rendering results of monocular single agent exploration on UT-MM and TUM dataset with Depth Pro as MMDE. We can see that the rendering provides good results, but also some small artifacts. Leveraging a MMDE allows to bypass some of the poorly detailed area of the depth ground truth from UT-MM.

Method	Config	Square-1	Square-2	Ego-Centric-1	Fast-Straight
Magic-SLAM	RGBD	6.24	X	4.78	5.72
MAC-Ego3D	RGBD	40.07	77.24	39.08	19.44
Ours	RGB+IMU	42.33	55.24	4.26	6.07
		Slow-Straight-1	Slow-Straight-2	Ego-Drive	Ego-Centric-2
Magic-SLAM	RGBD	8.49	5.94	8.01	7.16
MAC-Ego3D	RGBD	13.82	16.02	X	19.31

Appendix

Table S4 : Single agent tracking evaluation ATE (cm). Ours (RGB) against Magic-SLAM (RGB-D) and MAC-Ego3D (RGB-D) on UT-MM scenes.

3DGS SLAM

3D Gaussian Splatting simultaneous localization and mapping builds renderable 3D Gaussian maps while estimating camera motion online.

See all 63 papers

MAGS-SLAM: Monocular Multi-Agent Gaussian Splatting SLAM for Geometrically and Photometrically Consistent Reconstruction
May 11, 2026Zhihao Cao, Qi Shao, Shuhao Zhai +4Simultaneous Localization and Mapping 3D Gaussian Splatting
ETH Zurich, Switzerland · University of Liverpool, United Kingdom · Harbin Engineering University, China +4
Compact Keyframe-Optimized Multi-Agent Gaussian Splatting SLAM
Apr 1, 2026Monica M. Q. Li, Pierre-Yves Lajoie, Jialiang Liu +1Simultaneous Localization and Mapping 3DGS SLAM
Department of Computer Engineering, Polytechnique Montreal, Montreal, QC, Canada · Department of Mathematics, University of Wisconsin, Madison, WI 53706, USA
GS3LAM: Gaussian Semantic Splatting SLAM
Mar 29, 2026Linfei Li, Lin Zhang, Zhong Wang +13D Semantic Mapping 3DGS SLAM
School of Software Engineering, Tongji University Shanghai, China · Department of Automation, Shanghai Jiaotong University Shanghai, China

May 11, 2026J/K move · Enter open · S save

MAGS-SLAM: Monocular Multi-Agent Gaussian Splatting SLAM for Geometrically and Photometrically Consistent Reconstruction

Zhihao Cao, Qi Shao, Shuhao Zhai, Jing Zhang+3

ETH Zurich, Switzerland · University of Liverpool, United Kingdom · Harbin Engineering University, China+4

Simultaneous Localization and Mapping 3D Gaussian Splatting Multi-Robot Systems 3DGS SLAM 3D Reconstruction Visual SLAM

Open paper AskPDF arXiv

Collaborative photorealistic 3D reconstruction from multiple agents enables rapid large-scale scene capture for virtual production and cooperative multi-robot exploration. While recent 3D Gaussian Splatting (3DGS) SLAM algorithms can generate high-fidelity real-time mapping, most of the existing multi-agent Gaussian SLAM methods still rely on RGB-D sensors to obtain metric depth and simplify cross-agent alignment, limiting their deployment on low-cost or power-constrained robotic platforms, especially given the wider availability of RGB cameras. To address this challenge, we propose MAGS-SLAM, the first RGB-only multi-agent 3DGS SLAM framework for collaborative scene reconstruction. Each agent independently builds local monocular Gaussian submaps and transmits compact submap summaries rather than raw observations. To facilitate robust collaboration in the presence of monocular scale ambiguity, our framework integrates compact submap communication, geometry- and appearance-aware loop verification, and occupancy-aware Gaussian fusion, enabling coherent global reconstruction without active depth sensors. We further introduce ReplicaMultiagent Plus, a benchmark containing larger robot teams for evaluating collaborative Gaussian SLAM. Extensive experiments on both synthetic and real-world datasets demonstrate that MAGS-SLAM, using RGB images alone, achieves tracking accuracy, rendering quality, and geometric quality that are competitive with or superior to those of state-of-the-art RGB-D collaborative Gaussian SLAM methods.

Abstract

3DGS SLAM

Explore similar work

MAGS-SLAM: Monocular Multi-Agent Gaussian Splatting SLAM for Geometrically and Photometrically Consistent Reconstruction

Compact Keyframe-Optimized Multi-Agent Gaussian Splatting SLAM

GS3LAM: Gaussian Semantic Splatting SLAM