RoMo-GS: Efficient Online 3D Gaussian Splatting for Robot Mapping via Exposure-Integrated Motion Blur

Sangwoo Jung; Di Deng; Jiahui Fu; Laura Herlant; Ayoung Kim; Yafei Hu

cs.ROOct 17, 2026

RoMo-GS: Efficient Online 3D Gaussian Splatting for Robot Mapping via Exposure-Integrated Motion Blur

Authors: Sangwoo Jung, Di Deng, Jiahui Fu, Laura Herlant, Ayoung Kim, Yafei Hu

Organizations: RAI Institute, United States · Department of Mechanical Engineering, Seoul National University, South Korea

Abstract

Online 3D Gaussian Splatting (3DGS) is emerging as a practical way to build a representation for robotics applications. However, deploying it on real-world platforms exposes multiple challenges. Rapid maneuvers of a robot induce motion-blurred images, which corrupt the reconstructed geometry and cause camera pose drift. Furthermore, the number of Gaussians grows without bound during continuous online mapping. In this study, we present RoMo-GS, an online 3DGS system designed to address these challenges. Instead of averaging N discrete sub-frames as in prior motion blur-aware 3DGS, we analytically integrate each projected Gaussian over the exposure interval, yielding an O(1) complexity per Gaussian. Building on this, we bound the Gaussians' growth with a Hessian-based pruning score, while jointly refining pose, velocity, and Gaussians with Adam and an online LM step over the Gaussians. On multiple public datasets and a self-collected dataset, RoMo-GS matches or exceeds offline deblurring 3DGS methods in rendering quality while running 15--150x faster, and operates online at competitive quality with about 20% fewer Gaussians than the online 3DGS baseline.

Figures & tables

Figure 1: We introduce RoMo-GS , an efficient online 3D Gaussian Splatting system for real-world robotic mapping. Given a stream of motion-blurred frames captured by robots and front-end vision tracker outputs (a), RoMo-GS produces a compact 3DGS map (b) and sharp renderings online (c).

Figure 2: Overview of RoMo-GS . Given inputs from the vision-based tracker, we spawn initial Gaussians with the landmarks and depth samples. The exposure-integrated motion blur model renders a blurry image in a single rasterization. A joint Adam loop and an LM step refine the camera state and Gaussians, while Hessian-based score prunes redundant Gaussians. The pipeline produces a compact 3DGS map and sharp online renderings.

Figure 3: Comparison with multi-sample. (a) Static Gaussian with a motion arrow. (b) $N_{\text{blur}}=5$ , visible discretization. (c) $N_{\text{blur}}=100$ , smoother, but expensive. (d) Our model results same as (c) at a single sample.

	BAD-Gaussians	Deblurring 3DGS	GSOM	BAGS	CoMoGaussian	Ours
PSNR $\boldsymbol{\uparrow}$	28.367	24.174	22.895	24.606	23.650	25.148
SSIM $\boldsymbol{\uparrow}$	0.903	0.768	0.775	0.767	0.748	0.799
LPIPS $\boldsymbol{\downarrow}$	0.055	0.097	0.274	0.090	0.102	0.179
Train Time (sec)	2073.03	623.92	1431.63	1938.75	6203.25	41.68
# of Gaussians	261,602	168,837	121,553	141,271	55,845	75,813

Table 1: Deblur-NeRF Synthetic Dataset Experiment Results

	Deblurring 3DGS	GSOM	BAGS	CoMoGaussian	Ours
PSNR $\boldsymbol{\uparrow}$	14.278	25.879	19.414	19.764	25.367
SSIM $\boldsymbol{\uparrow}$	0.545	0.881	0.842	0.842	0.905
LPIPS $\boldsymbol{\downarrow}$	0.430	0.186	0.202	0.147	0.113
Train Time (sec)	685.73	1648.60	2922.50	7518.50	117.74
# of Gaussians	141,926	112,874	153,041	54,777	117,669

Table 2: ETH3D Real World Dataset Experiment Results

Figure 4: Top: ETH3D real-world. Bottom: Deblur-NeRF synthetic. GSOM fails to handle motion blur on both. Deblurring 3DGS and BAGS handle the synthetic case but collapse on ETH3D. CoMoGaussian is competitive on both, while Ours matches CoMoGaussian’s visual quality and trains 70–150 $\times$ faster online.

	RTG-SLAM	GS-ICP	MonoGS	Photo-SLAM	HI-SLAM2	tttLRM	MODP	Ours
PSNR $\boldsymbol{\uparrow}$	16.794	20.600	24.247	17.187	21.811	17.954	25.267	25.284
SSIM $\boldsymbol{\uparrow}$	0.617	0.768	0.803	0.652	0.784	0.714	0.842	0.840
LPIPS $\boldsymbol{\downarrow}$	0.461	0.220	0.227	0.334	0.200	0.363	0.187	0.194
Running Speed (fps)	7.540	30.143	2.420	46.585	4.827	12.239	11.335	11.044
# of Gaussians	303,425	1,584,402	37,811	66,395	93,609	3,726,362	210,275	167,234

Table 3: TUM Real World Dataset Experiment Results

	RTG-SLAM	GS-ICP	MonoGS	Photo-SLAM	HI-SLAM2	tttLRM	MODP	Ours
PSNR $\boldsymbol{\uparrow}$	20.460	18.503	23.387	21.254	23.640	21.316	26.994	26.822
SSIM $\boldsymbol{\uparrow}$	0.718	0.740	0.793	0.757	0.832	0.748	0.871	0.867
LPIPS $\boldsymbol{\downarrow}$	0.512	0.380	0.398	0.441	0.276	0.382	0.202	0.213
Running Speed (fps)	7.275	29.081	1.424	53.744	4.066	13.822	9.091	9.150
# of Gaussians	405,179	3,279,379	81,868	268,928	117,082	3,725,979	385,557	309,111

Table 4: Custom Real World Dataset Experiment Results

Figure 5: Scene reconstruction result of RoMo-GS on custom dataset. The red line shows the robot trajectory, including jittering due to the system’s physical interaction with the environment, and the yellow squares are the camera frustums.

	Ours (Exposure-Integrated)	w/o Blur Reduction	Multi Sample ( $N_{\text{blur}}=5$ )	Multi Sample ( $N_{\text{blur}}=10$ )
PSNR $\boldsymbol{\uparrow}$	25.367	24.355	25.951	26.003
SSIM $\boldsymbol{\uparrow}$	0.905	0.892	0.911	0.911
LPIPS $\boldsymbol{\downarrow}$	0.113	0.202	0.121	0.107
Running Speed (fps)	3.695	5.698	2.448	1.592
# of Gaussians	117,669	129,128	124,831	126,902

Table 5: Comparison with Multi-Sample Rendering on ETH3D Dataset

Figure 6: Comparison of exposure-integrated blur with multi-sample blur. $N_{\text{blur}}=5$ partially reduces motion blur. Compared with $N_{\text{blur}}=10$ , our model provides a competitive result with $2\times$ running speed.

Figure 7: Gaussian number growth on dog_park sequence. The number of Gaussians converges with Hessian-based Online Gaussian Pruning, while the Gaussian number of others grows unbounded or fluctuates unstably.

	PSNR	SSIM	LPIPS	# of Gaussians
Ours	25.148	0.799	0.179	75,813
Ours (150K Gaussians)	25.499	0.815	0.164	150,164
Deblurring 3DGS	24.174	0.768	0.097	168,837

Table 7: Experiment with

\sim

150K Gaussians on the Deblur-NeRF Synthetic Dataset

	PSNR	SSIM	LPIPS	# of Gaussians
Ours	25.648	0.908	0.113	106,285
MODP	23.726	0.879	0.244	129,165

Table 8: Comparison with MODP on the camera_shake_1 sequence of ETH3D

Figure 8: RoMo-GS and MODP on the blurry input (camera_shake_1 seq of ETH3D).

Pruning	Deblurring	LM	Pose/velocity	PSNR	SSIM	LPIPS	# of Gaussians
✓	✓	✓	✓	24.145	0.876	0.232	62,582
✗	✓	✓	✓	25.648	0.908	0.113	106,285
✗	✗	✓	✓	23.485	0.876	0.247	139,051
✗	✗	✗	✓	23.457	0.876	0.250	131,815
✗	✗	✗	✗	23.059	0.872	0.251	122,100

Table 9: Complete system ablation study of RoMo-GS on the camera_shake_1 sequence of ETH3D

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Real-world robot platforms used to acquire the custom dataset. (a) Handheld system with an Intel RealSense D435i RGB-D camera. (b) Spot quadruped robot with an Intel RealSense D455 RGB-D camera. (c) Franka robot arm with an Intel RealSense D405 RGB-D camera.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,013 papers

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM
Aug 14, 2026Thai Luu, Quan Tran, Hieu Phan +13DGS SLAM 3D Gaussian Splatting
Cognitive Robotics Lab, Department of Electrical Engineering and Computer Science, University of Arkansas, USA. · Center for AI Research, VinUniversity, Ha Noi, Vietnam.
GaussLite: Online Task-Conditioned 3D Gaussian Splatting for Real-Time Robotic Mapping
Jun 29, 2026Annika Thomas, Mason Peterson, Jonathan P. HowSimultaneous Localization and Mapping 3D Gaussian Splatting
MoonSplat: Monocular Online Gaussian Splatting with Sim(3) Global Optimization
Jun 16, 2026Guo Pu, Yixuan Han, Haofeng Li +33D Gaussian Splatting Language-Embedded 3DGS
Wangxuan Institute of Computer Technology, Peking University, China · Beijing Hydrogen Intelligent Tech. Co., Ltd., China

Aug 14, 2026J/K move · Enter open · S save

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM

Thai Luu, Quan Tran, Hieu Phan, Tuan Dang

Cognitive Robotics Lab, Department of Electrical Engineering and Computer Science, University of Arkansas, USA. · Center for AI Research, VinUniversity, Ha Noi, Vietnam.

3DGS SLAM 3D Gaussian Splatting Adaptive Gaussian Densification Streaming 3D Reconstruction

Open paper AskPDF arXiv

Recent 3D Gaussian Splatting (3DGS) has enabled efficient photorealistic view synthesis and is rapidly being adopted in simultaneous localization and mapping (SLAM) systems for online mapping. In these systems, a Gaussian map must be expanded and refined incrementally while tracking runs in real time, so initialization and density control directly determine where limited computation and iterations are spent. This contrasts with offline 3DGS reconstruction, where such heuristics can be amortized over long optimization schedules. However, most 3DGS-SLAM pipelines inherit initialization and density-control heuristics from offline reconstruction, which can become brittle under the strict per-keyframe optimization budgets and incremental map growth of online SLAM. In this work, we revisit these heuristics in a decoupled 3DGS-SLAM setting and propose three geometry-aware methods that operate in the mapping thread: transmittance-preserving densification, camera-aware scale initialization from depth and intrinsics, and error-guided densification that focuses new primitives on high-residual regions. Our results show consistent improvements in rendering quality with negligible overhead, highlighting the coupling between photometric residuals and pose uncertainty in online SLAM. We will open-source our code to the community to foster growth and validate reproducibility.

Abstract

3D Gaussian Splatting

Explore similar work

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM

GaussLite: Online Task-Conditioned 3D Gaussian Splatting for Real-Time Robotic Mapping

MoonSplat: Monocular Online Gaussian Splatting with Sim(3) Global Optimization