(b) Rendering w/o PD model(c) Rendering w/ PD model
Co-GS SLAM on our in-house dataset. (a) The merged
map of two agents, with their keyframe poses drawn as camera frusta.
(b) Without the pose-dependent (PD) Gaussian model, the misalignment
left by the merge is absorbed into the map as double surfaces and
blur. (c) With the PD model, the poses move together with the map
and the surface is recovered. Both renders start from the same merge
and differ only in the map representation.
8/9
Sequences with the lowest trajectory error
+5.9–10.5 dB
PSNR gain over the best baseline on all seven keyframe benchmarks
4.1 cm
Average in-house ATE, against 11.8 cm
The idea
Consistency starts with the poses.
Abstract
Collaborative 3D Gaussian splatting (3DGS) simultaneous localization
and mapping (SLAM) provides scalable intelligence by merging the
Gaussian maps that multiple agents build. To bring those maps into a
common frame, each agent tracks its own odometry, and the server
then closes loops between the agents by recognizing revisited places
and estimating the relative pose of the matched keyframes. Building
on this, state-of-the-art multi-agent 3DGS SLAM methods reconstruct
scenes with remarkable rendering quality. However, once they are
deployed in the real-world, they often suffer from two challenges:
(i) intra-agent odometry error, where the odometry of each
agent is not consistently accurate, and (ii)
inter-agent constraint noise, where sensor noise corrupts
the loop closures between the agents. To address these challenges,
we present Co-GS SLAM, which refines the merged map and the
keyframe poses of every agent together using a pose-dependent (PD)
Gaussian model, achieving a geometrically and photometrically
consistent merged map. Experiments on nine multi-agent sequences
from four datasets demonstrate the highest rendering fidelity and
the lowest trajectory error on all but one sequence. In addition, on
our in-house dataset, recorded in the real-world, our Co-GS SLAM
shows robust performance compared with state-of-the-art methods,
averaging 4.1 cm of trajectory error against 11.8 cm and 10.0 dB
more PSNR. Our code will be available upon acceptance.
How it works
From local maps to one shared map.
The server merges the agents’ maps with no prior on their relative
pose, then refines the map and every keyframe pose together.
Each agent tracks its own RGB-D stream and maps it into Gaussian
submaps in its own reference frame (a). The server detects the
loops between agents and solves a pose graph over the submaps,
which places them in a common frame as one rigid transform each
(b). We then refine the merged map together with the keyframe
poses of every agent, with each Gaussian defined in the camera
frame of the keyframe that created it (c).
01 / AGENT
Map in its own frame
Each agent estimates its keyframe poses by frame-to-model G-ICP
and builds Gaussian submaps in its own reference frame. Its
odometry error accumulates as it moves.
02 / SERVER
Merge into a common frame
Loop closures between agents enter a pose graph over all
submaps. Optimizing it moves each submap rigidly, so errors
inside a submap and noisy loop constraints remain after the
merge.
03 / SERVER
Refine map and poses together
The rendering loss over the keyframes of all agents is minimized
over the Gaussians and the keyframe poses of every agent at
once, starting from the merged map.
The key representation
Gaussians that move with their anchor poses.
Freeing the poses is not enough: from a poor merge, the joint
refinement can settle with the region two agents share
reconstructed twice.
The pose-dependent (PD) Gaussian model defines
each Gaussian in the camera frame of the keyframe that added it,
its anchor. A render then constrains not only its own
viewing pose but also the anchor poses of every Gaussian it
draws. Some of those anchors belong to the other agent, so the
agents correct each other where their views overlap.
w/o PD
7.31 cm ATE
w/ PD
3.38 cm ATE
(a) w/o PD(b) w/ PD(c) Ground truth
Aria room 0 after the joint refinement, with noise
injected into the inter-agent loop constraints. Keeping the
Gaussians in the world frame (a) leaves a second contour along
the back of the red chair and the arm of the black chair; the PD
model (b) recovers a single surface. The two renders differ by
only 0.15 dB PSNR, so this misalignment is one that PSNR barely
reports.
Quantitative results
Sharper maps. More accurate trajectories.
Nine multi-agent sequences from four datasets, evaluated on the
keyframes of the merged map.
Linear scale from zero. A bar with a break runs past the axis; its
exact value is printed.
All results · exact values from the paper
Table I · Tracking accuracy
Table I: tracking accuracy (ATE RMSE) on nine sequences
Method
Replica
Aria
TUM
In-house
apt 0
apt 1
apt 2
office 0
room 0
room 1
desk
Dream 1
Dream 2
ATE RMSE ↓ (cm)
MAGiC-SLAM
0.240
0.382
0.482
0.285
1.751
0.737
14.30
7.9
15.7
MAC-Ego3D*
286.0
396.1
390.3
217.4
224.9
166.7
15.0
52.5
156.5
CoKo-SLAM
0.62
0.56
0.35
0.55
2.77
1.62
4815
15.4
38.8
Ours (map only)
0.201
0.262
0.154
0.162
0.799
1.380
4.240
5.211
6.711
Co-GS SLAM (w/ PD)
0.190
0.248
0.148
0.122
0.805
1.362
2.114
3.394
4.783
Table II · Rendering quality and depth accuracy
Table II: rendering quality and depth accuracy on keyframes
Method
Replica
TUM
In-house
apt 0
apt 1
apt 2
office 0
desk
Dream 1
Dream 2
PSNR ↑ (dB)
MAGiC-SLAM
37.53
31.31
31.35
39.46
15.77
23.56
21.56
MAC-Ego3D*
31.16
22.30
29.23
27.58
17.19
25.15
24.40
CoKo-SLAM
36.65
29.25
31.27
39.27
14.81
23.11
20.98
Ours (map only)
46.32
38.49
40.72
44.63
17.23
28.44
29.25
Co-GS SLAM (w/ PD)
46.89
38.99
41.36
45.34
23.39
34.55
34.93
SSIM ↑
MAGiC-SLAM
0.982
0.963
0.961
0.992
0.615
0.846
0.772
MAC-Ego3D*
0.942
0.855
0.928
0.916
0.722
0.896
0.897
CoKo-SLAM
0.977
0.945
0.959
0.991
0.522
0.821
0.742
Ours (map only)
0.991
0.981
0.984
0.990
0.667
0.907
0.912
Co-GS SLAM (w/ PD)
0.992
0.982
0.986
0.991
0.826
0.961
0.960
LPIPS ↓
MAGiC-SLAM
0.098
0.148
0.179
0.058
0.517
0.342
0.415
MAC-Ego3D*
0.176
0.287
0.180
0.191
0.337
0.115
0.119
CoKo-SLAM
0.112
0.187
0.185
0.062
0.611
0.347
0.437
Ours (map only)
0.024
0.054
0.049
0.026
0.437
0.235
0.237
Co-GS SLAM (w/ PD)
0.022
0.053
0.047
0.024
0.259
0.147
0.152
Depth L1 ↓ (cm)
MAGiC-SLAM
0.63
1.57
0.98
0.39
9.40
11.47
25.72
MAC-Ego3D*
9.04
29.02
18.98
22.97
6.10
10.56
16.14
CoKo-SLAM
0.73
2.15
1.31
0.41
9.50
13.56
28.99
Ours (map only)
0.22
0.79
0.43
0.27
4.42
1.83
3.00
Co-GS SLAM (w/ PD)
0.20
0.75
0.40
0.24
5.85
2.42
4.15
Table III · Novel and training view synthesis on Aria
Table III: novel and training view synthesis on AriaMultiagent
Method
Novel views
Training views
room 0
room 1
Avg.
room 0
room 1
Avg.
PSNR ↑ (dB)
MAGiC-SLAM
22.16
22.08
22.12
26.85
27.44
27.15
MAC-Ego3D*
13.37
16.24
14.81
22.49
22.76
22.63
CoKo-SLAM
20.98
21.34
21.16
27.57
27.67
27.62
Co-GS SLAM (w/ PD)
25.97
22.39
24.18
30.41
31.97
31.19
SSIM ↑
MAGiC-SLAM
0.767
0.797
0.782
0.923
0.936
0.930
MAC-Ego3D*
0.527
0.693
0.610
0.807
0.839
0.823
CoKo-SLAM
0.740
0.770
0.755
0.933
0.935
0.934
Co-GS SLAM (w/ PD)
0.861
0.774
0.817
0.920
0.949
0.934
LPIPS ↓
MAGiC-SLAM
0.391
0.323
0.357
0.203
0.157
0.180
MAC-Ego3D*
0.592
0.566
0.579
0.359
0.446
0.403
CoKo-SLAM
0.407
0.344
0.376
0.201
0.164
0.183
Co-GS SLAM (w/ PD)
0.245
0.281
0.263
0.176
0.136
0.156
Depth L1 ↓ (cm)
MAGiC-SLAM
3.20
5.10
4.15
1.92
1.30
1.61
MAC-Ego3D*
90.81
33.78
62.30
36.22
23.41
29.82
CoKo-SLAM
3.48
6.12
4.80
1.43
1.66
1.55
Co-GS SLAM (w/ PD)
3.26
5.65
4.46
3.97
3.04
3.51
Bold marks the best result as ranked in the paper.
“Map only” is our refinement with the keyframe poses held fixed.
All methods run with the authors’ code under an unknown relative
pose, with the transform between agents initialized to the
identity. *MAC-Ego3D assumes that transform to be known, so it is
reported outside the setting it was designed for.
Qualitative results
One surface, seen from both agents.
The merged map rendered at a keyframe of each agent, with no
initial transform between the agents.
In-house · Dream 1
Agent 1MAGiC-SLAMMAC-Ego3D*CoKo-SLAMCo-GS SLAMGround truth
Agent 2MAGiC-SLAMMAC-Ego3D*CoKo-SLAMCo-GS SLAMGround truth
The two rows are the two overlapping viewpoints whose disagreement
the refinement uses. The baselines absorb the residual of the
merge as blur and duplicated surfaces over the desk and the chair,
while refining the poses together with the map recovers the
keyboard, the monitor and the chair frame from both sides.
TUM desk
Agent 1MAGiC-SLAMMAC-Ego3D*CoKo-SLAMCo-GS SLAMGround truth
Agent 2MAGiC-SLAMMAC-Ego3D*CoKo-SLAMCo-GS SLAMGround truth
Each TUM agent drifts on its own, so this is the hardest case of
intra-agent odometry error. The baselines absorb the pose error
as blur and smeared surfaces, while refining the poses together
with the map recovers the monitor, the keyboard and the books from
both viewpoints.
Real-world evaluation
Our in-house multi-agent sequences.
Two indoor sequences, each pairing two independent recordings as
two agents.
Sequences
Dream 1 · Dream 2
Agents
2 per sequence, 460–1103 frames each
Sensor
Intel RealSense D455f RGB-D
Ground truth
NOKOV motion capture
Citation
BibTeX
@inproceedings{sim2026cogsslam,
title = {Co-GS SLAM: Collaborative 3D Gaussian Splatting SLAM with Pose-Dependent Gaussian Model},
author = {Sim, Taehu and Kim, Hogyun and Choi, Jiwon and Lee, Juhui and Cho, Younggun},
booktitle = {IROS 2026 Workshop and Competition on Intelligent Information Gathering for Single and Multi-Robot Systems},
year = {2026}
}
Co-GS SLAM
