CoSphere-SLAM: Multi-Agent Panoramic Gaussian Splatting SLAM with Spherical Loop Closure
Authors: Qi Shao, Zhihao Cao, Shuhao Zhai, Taoyu Wu, Zhuoxiao Li, Anh Nguyen, Long Hu, Xianzhi Li, Hong Liang, Baoru Huang, Liang Li
Organizations: Harbin Engineering University, Harbin, China · the University of Liverpool, Liverpool, UK · ETH Zurich, Zurich, Switzerland · the National University of Singapore, Singapore · the University of Macau, Macau, China · Xi’an Jiaotong-Liverpool University, Suzhou, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Imperial College London, London, UK · Huazhong University of Science and Technology, Wuhan, China
Dense collaborative mapping is essential for search and rescue, building inspection, and warehouse automation, where panoramic cameras capture the entire surroundings in a single frame. Unlike pinhole agents that share views only at similar headings, panoramic agents observe overlapping scenes at any 3D orientation, yet searching full rotations also produces mutually consistent false loop closures across distinct symmetric rooms. Moreover, continuous monocular scale drift splits rigidly merged maps at revisited surfaces, while depth noise at short baselines distorts loop translation directions. To address these challenges, we propose CoSphere-SLAM, the first multi-agent Gaussian splatting SLAM framework for panoramic cameras, whose spherical loop closure aligns spherical features over SO(3) via fast Fourier correlation across arbitrary orientations and rejects symmetric false closures through global Sim(3) consensus. Building on these closures, directional global optimization solves a per-keyframe Sim(3) pose graph with log-scale regularization to correct scale drift while separating optical-flow translation directions from depth-based magnitudes. Finally, omnidirectional Gaussian fusion transforms each Gaussian with its keyframe similarity and renders Gaussians at their maximum ray response to jointly refine poses and maps. Experiments on indoor and cross-session multi-agent panoramic benchmarks demonstrate that CoSphere-SLAM achieves state-of-the-art trajectory accuracy and rendering quality among single-agent and multi-agent methods. Project website: https://cosphere-slam.github.io.
Figures & tables
Fig. 2: Overview of CoSphere-SLAM. Each agent resamples panoramas into six cube-map faces with scale-aligned depth to track poses and build local 3D Gaussians. On the server, spherical loop closure aligns spherical feature maps over SO(3) via circular FFT and removes symmetric false closures via global Sim(3) consensus across clusters and agent loops. Directional global optimization solves a per-keyframe Sim(3) pose graph separating optical-flow directions from depth-aligned magnitudes, and omnidirectional Gaussian fusion deforms Gaussians by their keyframe similarities to refine all poses and maps along viewing rays.
Development Scenes
Five Longest Test Scenes
All Test Scenes
Method
002
005
006
007
009
Mean
055
011
047
003
043
ATE
Track
Single-Agent, Front Cube Face
DROID-SLAM [ 32 ]
1.011
2.050
2.548
0.843
1.934
1.677
4.891
7.389
1.162
2.972
1.840
1.610 †
99%
DPVO [ 35 ]
1.836
2.373
2.601
1.159
1.802
1.954
5.993
6.084
1.289
2.622
1.741
1.576
100%
MASt3R-SLAM [ 36 ]
0.771
0.236
0.232 †
0.915
0.638
0.559 †
3.671 †
4.080
0.215
1.457
0.152
0.651 †
88%
TTT3R [ 26 ]
3.037
2.002
1.726
1.349
1.935
2.010
10.749
9.196
0.933
4.455
1.985
2.542 †
99%
TABLE I: Per-agent ATE (m) on the Holo360D multi-agent panoramic benchmark, grouped by front cube-face or full-panorama input. Mean averages the five development scenes, and All Test Scenes reports mean ATE and tracked rate over 45 test scenes, where † averages tracked agents and – marks untracked scenes. Dark and light orange mark 1st and 2nd among multi-agent methods.
Fig. 3: Omnidirectional Gaussian fusion. Left: evaluating Gaussians at their maximum ray response and compositing front to back avoids polar stretch and dashed seam cuts. Right: per-keyframe Sim(3) scaling forms the merged map while random 3D rotations move poles onto the equator for 2D SSIM supervision.
Fig. 4: Rendering comparison on four development and two test scenes. Each method renders a shared frame from its map at its estimated pose, showing the panorama on Scene 002 for panoramic methods, two cube faces for pinhole methods, and the front face elsewhere. Crosses mark failed runs, and black regions lie outside submaps.
ATE
Joint
Method
Scenes
Baseline
Ours
Baseline
Ours
Single-Agent
DROID-SLAM
44
1.620
0.149
–
–
DPVO
45
1.576
0.150
–
–
MASt3R-SLAM
31
0.677
0.169
–
–
TTT3R
44
2.585
0.152
–
–
TABLE II: Test trajectory comparison (ATE in m) after map refinement where each baseline tracks all agents. Single- and multi-agent baselines pair with Ours (Single) and Ours (Multi), Scenes counts valid scenes, and bottom rows cover all 45 scenes.
PSNR
LPIPS
Method
Scenes
Baseline
Ours
Baseline
Ours
ODGS-SLAM
43
17.67
27.92
0.620
0.231
CP-SLAM
19
20.85
27.60
0.602
0.241
MAC-Ego3D
12
16.32
29.36
0.658
0.214
MAGiC-SLAM
21
13.92
25.61
0.594
0.244
MAGS-SLAM
40
26.09
28.25
0.318
0.235
TABLE III: Rendering quality of the front cube face on test scenes where each baseline outputs full or partial submaps, paired with our maps on shared frames over all pixels without masking.
Method
Track
ATE
Novel WS
SSIM
Merged WS
PFGS360 [ 29 ]
80%
0.212 †
21.20
0.650
–
Ours (Single)
100%
0.139
28.54
0.736
24.28
w/o Global Map Refinement
100%
0.055
18.23
0.598
17.81
Ours (Multi)
100%
0.021
28.95
0.745
27.80
TABLE IV: Omnidirectional 3DGS evaluation at 2048×1024 on the development set with operator masking: tracked agents, per-agent ATE (m), Scene 009 novel-view WS-PSNR and SSIM, and best single-agent or merged full-scene WS-PSNR over five scenes.
Method
Scene 002
Scene 005
Scene 007
Total
ODGS+RICP [ 9 ]
3,869
5,454
5,809
15,132
Ours (Multi)
1,260
867
986
3,113
TABLE V: Multi-agent alignment wall-clock time (s) on Holo360D.
Joint ATE (m)
Method
ATE
Atrium
Conc.
Hall
Piat.
Track
Ours (Single)
0.88
–
–
–
–
100%
MAC-Ego3D [ 16 ]
23.59
20.69
28.30
29.43
34.07
100%
MAGS-SLAM [ 17 ]
2.04 †
14.41
20.10
–
34.62
83%
ODGS+RICP [ 9 ]
16.39 †
–
–
–
33.66
67%
OpenVSLAM-M [ 27 ]
0.38 †
–
0.14
–
–
72%
TABLE VI: Cross-session mapping ATE (m) over 18 day and night handheld sessions on 360Loc [ 40 ] ( † partial, – unmerged).
Development
Test
Variant
ATE
Joint
ATE
Joint
Ours (Single)
0.139
–
0.150
–
w/ MAGS-SLAM Association
0.172
1.329
–
–
Pinhole Front-Face Only
0.075
0.221
0.225
0.561
w/o Cycle Consistency
0.022
0.029
0.098
0.355
w/o Per-Keyframe Scale
0.092
0.162
0.144
0.245
TABLE VII: Ablation on shared six-face inputs (ATE / joint ATE in m, † 3 scenes). Development and final test rows include refinement, while other test rows follow the pose graph.
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.
Zhihao Cao, Qi Shao, Shuhao Zhai +4
ETH Zurich, Switzerland · University of Liverpool, United Kingdom · Harbin Engineering University, China +4
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
Monica M. Q. Li, Pierre-Yves Lajoie, Jialiang Liu +1
Department of Computer Engineering, Polytechnique Montreal, Montreal, QC, Canada · Department of Mathematics, University of Wisconsin, Madison, WI 53706, USA
Outdoor robot teams need a shared dense map despite limited overlap, independent reference frames, and uncertain monocular scale. Collaborative dense SLAM systems typically resolve this with depth sensors, which add payload, power, and calibration cost. We present CoMo3R-SLAM, a collaborative monocular dense SLAM system that places learned feed-forward 3D reconstruction priors at the center of the multi-agent problem: their dense pointmaps anchor scale across agents and supply correspondences strong enough to verify inter-agent links geometrically. Each agent tracks and fuses its own keyframes from a single RGB stream, while a coordinator retrieves cross-agent keyframes over the prior's encoder features, verifies them by bidirectional dense pointmap matching, synchronizes the independent similarity gauges in closed form, and refines every keyframe in one unified multi-agent sim(3) graph. Finally, a pose-depth alternation over geometry-aware segments lets inter-agent observations constrain dense structure as well as trajectories. Requiring neither measured depth nor supplied intrinsics, CoMo3R-SLAM attains the lowest trajectory error on three of four Tanks and Temples scenes, and competitive accuracy on Waymo driving sequences, while running at approximately 6-8 FPS on RTX 3080 Ti. A long-horizon traversal, independently captured day and night streams, and teams of up to four agents further map its operating range.
Zhihao Cao, Qi Shao, Shuhao Zhai +4
ETH Zurich, Switzerland. · Harbin Engineering University, China. · University of Liverpool, United Kingdom. +4