Compact Keyframe-Optimized Multi-Agent Gaussian Splatting SLAM

Monica M. Q. Li; Pierre-Yves Lajoie; Jialiang Liu; Giovanni Beltrame

cs.ROApr 1, 2026

Compact Keyframe-Optimized Multi-Agent Gaussian Splatting SLAM

Authors: Monica M. Q. Li, Pierre-Yves Lajoie, Jialiang Liu, Giovanni Beltrame

Organizations: Department of Computer Engineering, Polytechnique Montreal, Montreal, QC, Canada · Department of Mathematics, University of Wisconsin, Madison, WI 53706, USA

Abstract

Efficient multi-agent 3D mapping is essential for robotic teams operating in unknown environments, but dense representations hinder real-time exchange over constrained communication links. In multi-agent Simultaneous Localization and Mapping (SLAM), systems typically rely on a centralized server to merge and optimize the local maps produced by individual agents. However, sharing these large map representations, particularly those generated by recent methods such as Gaussian Splatting, becomes a bottleneck in real-world scenarios with limited bandwidth. We present an improved multi-agent RGB-D Gaussian Splatting SLAM framework that reduces communication load while preserving map fidelity. First, we incorporate a compaction step into our SLAM system to remove redundant 3D Gaussians, without degrading the rendering quality. Second, our approach performs centralized loop closure computation without initial guess, operating in two modes: a pure rendered-depth mode that requires no data beyond the 3D Gaussians, and a camera-depth mode that includes lightweight depth images for improved registration accuracy and additional Gaussian pruning. Evaluation on both synthetic and real-world datasets shows up to 85-95% reduction in transmitted data compared to state-of-the-art approaches in both modes, bringing 3D Gaussian multi-agent SLAM closer to practical deployment in real-world scenarios. Code: https://github.com/lemonci/coko-slam

Figures & tables

Panel of Figure 1 — Figure 1: Top view comparison of merged maps of two agents in scene office-0 of the Replica dataset [ 17 ] . Our method, Coko-SLAM, merges the submaps from different agents correctly while CP-SLAM [ 19 ] (a), MAGiC-SLAM [ 20 ] (b), and MAC-Ego3D [ 23 ] (c) fail without initial relative poses between agents. The maps in (a) and (d) are merged using rendered images from the submaps; the maps in (b), (c), and (e) are merged using depth images directly from the camera.

Figure 2: We adapt a keyframing method to select keyframes based on distance in feature space. We then eliminate redundant 3D Gaussians during mapping. The agents then send pure 3D Gaussian submaps with keyframe feature vectors to a server for loop closure and map merging, without the initial relative poses among agents.

Panel of Figure 3 — Algorithm 1 Procedure of Optimization-Sparsification

Methods	Metrics	Office-0	Apart-0	Apart-1	Apart-2
CP-SLAM	PSNR [dB] $\uparrow$	24.257 (2.224)	24.56 (2.502)	19.185 (3.800)	25.425 (4.209)
	SSIM $\uparrow$	0.749 (0.104)	0.765 (0.099)	0.646 (0.177)	0.82 (0.173)
	LPIPS $\downarrow$	0.434 (0.130)	0.437 (0.139)	0.555 (0.176)	0.403 (0.194)
	Depth L1 [cm] $\downarrow$	12.637 (12.727)	14.618 (14.746)	26.619 (27.384)	5.061 (5.222)
MAGiC-SLAM	PSNR [dB] $\uparrow$	25.407 (0.237)	✗	✗	✗
	SSIM $\uparrow$	0.295 (0.005)	✗	✗	✗

Table 1: Training view synthesis performance on Replica Multiagent dataset without initial relevant poses between agents. The merged map based on multiple agents’ submaps is evaluated by synthesizing training views. ✗ indicates failures during map merging. Bold = best, underlined = second best.

Methods	Metrics	Novel Views		Training Views
Methods	Metrics	Room0	Room1	Room0	Room1
CP-SLAM	PSNR [dB] $\uparrow$	8.572 (0.286)	9.571 (0.368)	9.384 (0.632)	9.063 (0.814)
	SSIM $\uparrow$	0.292 (0.02)	0.268 (0.033)	0.94 (0.044)	0.959 (0.038)
	LPIPS $\downarrow$	0.879 (0.01)	0.914 (0.028)	0.208 (0.029)	0.257 (0.046)
	Depth L1 [cm] $\downarrow$	177.519 (10.136)	123.016 (7.009)	198.469 (197.717)	138.848 (138.472)
MAGiC-SLAM	PSNR [dB] $\uparrow$	13.506 (0.125)	19.092 (0.332)	20.24 (0.074)	22.772 (0.107)

Table 2: Novel view synthesis synthesis performance on Aria Multiagent dataset without initial relevant poses between agents. The merged map based on multiple agents’ submaps is evaluated by synthesizing novel views. The merged map based on multiple agents’ submaps is evaluated by synthesizing novel and training views. Bold = best, underlined = second best.

	Replica								Aria
	Office-0		Apart-0		Apart-1		Apart-2		Room0			Room1
Methods	A0	A1	A0	A1	A0	A1	A0	A1	A0	A1	A2	A0	A1	A2
CP-SLAM	150	150	250	250	250	250	250	250	50	50	-	50	50	-
MAGiC-SLAM	300	300	500	500	500	500	500	500	100	100	100	100	100	100
MAC-Ego3D	150	150	250	250	250	250	250	250	50	50	50	50	50	50
Ours	160	147	196	235	170	171	194	176	84	89	94	86	89	92

Table 3: Number of keyframes per agent across different scenes. Comparison for Replica and Aria datasets. - indicates that CP-SLAM does not support more than two agents. Bold = best (fewest), underlined = second best.

	Replica								Aria
	Office-0		Apart-0		Apart-1		Apart-2		Room0			Room1
Methods	A0	A1	A0	A1	A0	A1	A0	A1	A0	A1	A2	A0	A1	A2
CP-SLAM	376	376	626	626	627	627	626	626	45	44	-	43	43	-
MAGiC-SLAM	2034	2070	3471	3474	3569	3543	3537	3549	538	531	515	541	530	512
MAC-Ego3D	2061	2008	5027	4861	6551	3404	3448	3329	656	2305	2391	554	631	410
Ours(Rendered depth)	98	108	187	230	247	256	194	202	61	113	79	89	68	56

Table 4: Total data transmitted in megabytes per agent across different scenes. Comparison for Replica and Aria datasets. - indicates that CP-SLAM does not support more than two agents. Bold = best (least), underlined = second best.

	Replica								Aria
	Office-0		Apart-0		Apart-1		Apart-2		Room0			Room1
Methods	A0	A1	A0	A1	A0	A1	A0	A1	A0	A1	A2	A0	A1	A2
Pcd	2034	2070	3471	3474	3569	3543	3537	3549	538	531	515	541	530	512
Pcd + C	2027	2057	3455	3452	3542	3522	3516	3533	525	516	505	529	520	504
Pcd + K	1471	1760	1353	1373	1317	1221	1568	1441	256	386	354	325	269	278
Pcd + K + C	1466	1750	1330	1344	1290	1204	1545	1419	238	366	338	306	255	265

Table 5: Total data transmitted in megabytes under different settings. Transmitted data is reported w.r.t. the combination of settings: (1) Data type: Point cloud ( Pcd , same as MAGiC-SLAM), rendered depth ( RD , pure 3D Gaussian map) and camera depth ( CD , 3D Gaussian map with depth image) (2) Keyframing ( K ) (3) Compaction ( C ). Bold = best (least), underlined = second best.

Simultaneous Localization and Mapping

Simultaneous localization and mapping estimates a robot’s motion while building a map of an unknown environment.

See all 197 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
CGS-SLAM: Collaborative Gaussian Splatting based SLAM for Multi-Agent Reconstruction
Aug 27, 2026Jean-Daniel de Ambrogi, Aladine Chetouani, Vincent Nguyen +13DGS SLAM Visual-Inertial SLAM
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
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

Simultaneous Localization and Mapping

Explore similar work

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

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

GS3LAM: Gaussian Semantic Splatting SLAM