F$^2$ SLAM: Turning Feed-Forward Geometry into Persistent Factors for SLAM

Zhisong Xu; Fan Zhu; Jiawei Qian; Ziyu Chen; Zhenjun Zhao; Javier Civera

cs.CVOct 4, 2026

F $^2$ SLAM: Turning Feed-Forward Geometry into Persistent Factors for SLAM

Authors: Zhisong Xu, Fan Zhu, Jiawei Qian, Ziyu Chen, Zhenjun Zhao, Javier Civera

Organizations: The University of Tokyo · Tohoku University · University of Science and Technology of China · University of Zaragoza

Abstract

Feed-forward 3D models provide strong multi-view geometric priors, while on- line simultaneous localization and mapping (SLAM) relies mainly on local mea- surements and can accumulate drift over long sequences. Existing attempts to combine the two typically treat feed-forward predictions as an external geomet- ric state that is aligned or fused with the online estimate after the fact, which keeps broader multi-view evidence outside the optimizer that refines the SLAM state. We present F2SLAM, which instead converts feed-forward geometry di- rectly into optimization-native target-weight measurements attached to a persis- tent dense factor graph. A high-frequency stream maintains local tracking con- straints and graph connectivity, while a low-frequency stream uses wider multi- view context to selectively refresh existing measurements after a state-consistency check. Both streams constrain the same poses, inverse depths, and optional cam- era intrinsics through a single dense bundle adjustment. Experiments on multiple benchmarks demonstrate consistently strong trajectory estimation and improved dense reconstruction in both calibrated and uncalibrated settings. Notably, the uncalibrated configuration reduces the average ATE RMSE from 0.030 m for the strongest feed-forward baseline to 0.002 m on the Replica dataset.

Figures & tables

Figure 1: Motivation and overview of F 2 SLAM. Conventional dense SLAM (top) relies on local frame-to-frame measurements, while foundation-model-based approaches (middle) reconstruct and align local submaps. In contrast, F 2 SLAM (bottom) integrates high-frequency tracking and low-frequency feed-forward geometry in a persistent dense factor graph, where low-frequency predictions selectively refresh existing measurements for joint refinement through a single dense BA. The inset reports ATE RMSE on the TUM RGB-D dataset Sturm et al. (2012) .

Figure 2: Pipeline of F 2 SLAM. The high-frequency stream selects keyframes, maintains the graph topology, and frequently updates local target-weight measurements. The low-frequency stream refreshes measurements on selected existing edges using wider multi-view context, without adding new edges. Both streams constrain the same persistent geometric state through a single dense BA.

Method	Uncalib.	360	desk	desk2	floor	plant	room	rpy	teddy	xyz	Avg.
DROID-SLAM ( Teed and Deng, 2021 )	✗	0.111	0.018	0.042	0.021	0.016	0.049	0.026	0.048	0.012	0.038
DPV-SLAM ( Lipson et al., 2024 )	✗	0.112	0.018	0.029	0.057	0.021	0.330	0.030	0.084	0.010	0.076
DPV-SLAM++ ( Lipson et al., 2024 )	✗	0.132	0.018	0.029	0.050	0.022	0.096	0.032	0.098	0.010	0.054
GO-SLAM ( Zhang et al., 2023 )	✗	0.089	0.016	0.028	0.025	0.026	0.052	0.019	0.048	0.010	0.035
MASt3R-SLAM ( Murai et al., 2025 )	✗	0.049	0.016	0.024	0.025	0.020	0.061	0.027	0.041	0.009	0.030
F 2 SLAM (ours)	✗	0.070	0.017	0.026	0.021	0.014	0.049	0.023	0.034	0.010	0.029

Table 1: Trajectory evaluation (ATE RMSE [m]

\downarrow

) on the TUM RGB-D dataset ( Sturm et al., 2012 ) .

*

denotes results reproduced on our hardware.

Method	Uncalib.	chess	fire	heads	office	pumpkin	kitchen	stairs	Avg.
DROID-SLAM ( Teed and Deng, 2021 )	✗	0.036	0.027	0.025	0.066	0.127	0.040	0.026	0.050
NICER-SLAM ( Zhu et al., 2024 )	✗	0.033	0.069	0.042	0.108	0.200	0.039	0.108	0.086
F 2 SLAM (ours)	✗	0.036	0.029	0.025	0.088	0.140	0.042	0.021	0.053
DROID-SLAM ( Teed and Deng, 2021 )	✓	0.047	0.038	0.034	0.136	0.166	0.080	0.044	0.078
MASt3R-SLAM ( Murai et al., 2025 )	✓	0.063	0.046	0.029	0.103	0.114	0.074	0.032	0.066
VGGT-SLAM ∗ ( Maggio et al., 2025 )	✓	0.037	0.026	0.018	0.104	0.133	0.061	0.093	0.067

Table 2: Trajectory evaluation (ATE RMSE [m]

\downarrow

) on the 7-Scenes dataset ( Shotton et al., 2013 ) .

*

denotes results reproduced on our hardware.

Method	Uncalib.	R0	R1	R2	O0	O1	O2	O3	O4	Avg.
DROID-SLAM ( Teed and Deng, 2021 )	✗	0.003	0.001	0.003	0.003	0.004	0.003	0.005	0.004	0.003
NICER-SLAM ( Zhu et al., 2024 )	✗	0.013	0.016	0.011	0.021	0.032	0.021	0.014	0.020	0.019
F 2 SLAM (Ours)	✗	0.003	0.002	0.003	0.002	0.003	0.003	0.003	0.004	0.003
VGGT-SLAM ( Maggio et al., 2025 )	✓	0.030	0.167	0.086	0.042	0.064	0.095	0.039	0.043	0.071
VGGT-SLAM 2.0 ∗ ( Maggio and Carlone, 2026 )	✓	0.030	0.049	0.040	0.026	0.018	0.024	0.024	0.029	0.030
SLAM-Former ∗ ( Yuan et al., 2026 )	✓	0.031	0.033	0.025	0.030	0.027	0.038	0.033	0.036	0.032

Table 3: Trajectory evaluation (ATE RMSE [m]

\downarrow

) on the Replica dataset ( Straub et al., 2019 ) .

*

denotes results reproduced on our hardware.

Method	Uncalib.	00	01	02	03	04	05	06	07	08	09	10	Avg.
ORB-SLAM2 ( Mur-Artal and Tardós, 2017 )	✗	40.65	502.20	47.82	0.94	1.30	29.95	40.82	16.04	43.09	38.77	5.42	69.73
LDSO ( Gao et al., 2018 )	✗	9.32	11.68	31.98	2.85	1.22	5.10	13.55	2.96	129.02	21.64	17.36	22.43
DROID-SLAM ( Teed and Deng, 2021 )	✗	92.10	5344.60	107.61	2.38	1.00	118.50	62.47	21.78	161.60	72.32	118.70	554.82
DPV-SLAM ( Lipson et al., 2024 )	✗	112.80	11.50	123.53	2.50	0.81	57.80	54.86	18.77	110.49	76.66	13.65	53.03
DPV-SLAM++ ( Lipson et al., 2024 )	✗	8.30	11.86	39.64	2.50	0.78	5.74	11.60	1.52	110.90	76.70	13.70	25.75
F 2 SLAM (ours)	✗	117.08	64.77	134.50	3.94	1.69	83.39	65.43	25.37	112.10	102.82	27.95	67.19

Table 4: Trajectory evaluation (ATE RMSE [m]

\downarrow

) on the KITTI Odometry dataset ( Geiger et al., 2012 ) . Baseline results are taken from VGGT-SLAM++ ( Mandal et al., 2026 ) .

	Room0		Room1		Room2		Office0		Office1		Office2		Office3		Office4		Avg.
Method	Acc.	Comp.	Acc.	Comp.	Acc.	Comp.	Acc.	Comp.	Acc.	Comp.	Acc.	Comp.	Acc.	Comp.	Acc.	Comp.	Acc.	Comp.
DROID-SLAM + ( Teed and Deng, 2021 )	0.1218	0.0896	0.0835	0.0607	0.0326	0.1601	0.0301	0.1619	0.0239	0.1619	0.0566	0.1556	0.0449	0.0973	0.0465	0.0963	0.0550	0.1229
NICER-SLAM + ( Zhu et al., 2024 )	0.0253	0.0304	0.0393	0.0410	0.0340	0.0342	0.0549	0.0609	0.0345	0.0442	0.0402	0.0429	0.0334	0.0403	0.0303	0.0387	0.0365	0.0416
MASt3R-SLAM ∗ ( Murai et al., 2025 )	0.2468	0.2675	0.0393	0.0266	0.0258	0.0187	0.0336	0.0211	0.0359	0.0303	0.0609	0.0484	0.0881	0.0824	0.0742	0.0577	0.0756	0.0691
SLAM-Former ∗ ( Yuan et al., 2026 )	0.0339	0.0247	0.0248	0.0171	0.0244	0.0166	0.0208	0.0153	0.0178	0.0138	0.0345	0.0222	0.0296	0.0213	0.0305	0.0214	0.0270	0.0187
VGGT-SLAM 2.0 ∗ ( Maggio and Carlone, 2026 )	0.0305	0.0622	0.0564	0.0765	0.0277	0.0629	0.0475	0.0457	0.0340	0.0595	0.0249	0.0792	0.0322	0.0739	0.0408	0.0743	0.0367	0.0657

Table 5: Reconstruction accuracy (Acc. [m]

\downarrow

) and completion (Comp. [m]

\downarrow

) on the Replica dataset ( Straub et al., 2019 ) .

+

denotes results reported in NICER-SLAM,

-

denotes results reported in SLAM3R, and

*

denotes results reproduced on our hardware.

Table 8

Figure 3: Qualitative reconstruction comparison on representative sequences. F 2 SLAM recovers more continuous geometry with fewer missing regions and artifacts.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Table 10

Method	Opt. Intr.	Chess	Fire	Heads	Office	Pumpkin	RedKitchen	Stairs	Avg.
w/ target motion 600	✓	0.036	0.024	0.014	0.092	0.143	0.048	0.025	0.054
w/ target motion 800 (Ours)	✓	0.036	0.024	0.014	0.087	0.140	0.041	0.020	0.051
w/ target motion 1000	✓	0.036	0.025	0.014	0.087	0.140	0.041	0.020	0.051

Appendix

Table 10: Ablation study of Flow3r target motion on 7-Scenes. We report Sim(3)-aligned ATE RMSE (m). Lower is better.

Figure 4: Ablation on the low-frequency context window size. Lower is better for all metrics. Log-scaled y-axes retain a continuous line plot while keeping the large errors at sizes 4 and 8 visible. We use a logarithmic coordinate system to better show the overall trend.

Figure 5: Qualitative trajectory comparison on representative sequences. All estimated trajectories are aligned to the ground truth using the same evaluation protocol. Compared with competing methods, F 2 SLAM follows the reference trajectory more closely throughout the sequence and exhibits less accumulated drift.

Figure 6: Two-dimensional trajectory projections on representative sequences. The projections emphasize global shape, turning behavior, and endpoint drift. F 2 SLAM better preserves the reference path and avoids the progressive deviation observed in competing methods.

Figure 7: Three-dimensional trajectory visualization. The full spatial trajectories reveal vertical drift and geometric distortion that can be hidden by planar projections. F 2 SLAM maintains a more coherent trajectory across all spatial directions.

Figure 8: Qualitative reconstruction results on representative Replica scenes. F 2 SLAM recovers coherent room layouts and preserves major scene structures across long sequences.

Figure 9: Qualitative reconstruction results on representative KITTI Odometry sequences. F 2 SLAM maintains improved global consistency and reduces accumulated drift over long driving sequences.

Explore similar work

UniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) Optimization
Aug 3, 2026Inha Lee, Dongjae Jeong, Junhee Lee +1Orb-Slam2 Simultaneous Localization And Mapping
lsan National Institute of Science and Technology, Ulsan, Korea
ScaRF-SLAM: Scale-Consistent Reconstruction with Feed-Forward Models and Classical Visual SLAM
May 29, 2026Yuhao Zhang, Yifu Tao, Frank Dellaert +1Orb-Slam2 Simultaneous Localization And Mapping
Oxford Robotics Institute, University of Oxford, UK · School of Interactive Computing, Georgia Institute of Technology, USA
GeoGS-SLAM: Online Monocular Reconstruction Using Gaussian Splatting with Geometric Priors
Jul 13, 2026Ruilan Gao, Letian Jin, Yu ZhangGaussian Splatting Simultaneous Localization And Mapping Systems Scene Reconstruction
State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou, China, 310027. · Key Laboratory of Collaborative Sensing and Autonomous Unmanned Systems of Zhejiang Province, Hangzhou, China, 310027.

Aug 3, 2026, cs.CVJ/K move · Enter open · S save

UniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) Optimization

Inha Lee, Dongjae Jeong, Junhee Lee, Kyungdon Joo

lsan National Institute of Science and Technology, Ulsan, Korea

Orb-Slam2 Simultaneous Localization And Mapping Multi-View Nvidia Isaac Sim

Open paper AskPDF arXiv

Recent geometric foundation models enable feed-forward inference for SLAM, but their predictions are strongly dependent on the input view set, which leads to geometric inconsistencies and trajectory drift when results are chained over long sequences. Online deployment further exposes a trade-off between the low latency of two-view tracking and the constraint richness of multi-view inference. We introduce UniSim-SLAM, an integrated system that runs lightweight two-view keyframe tracking in the frontend and performs periodic multi-view submap refinement in the backend. To combine predictions defined in heterogeneous local coordinates with inconsistent scales, we formulate a unified multi-level factor graph on

Sim(3)

that jointly optimizes global keyframe poses and submap poses. The graph integrates temporal view-to-view odometry edges, view-to-submap bridge edges with depth-statistics scale anchoring, and submap-to-submap tie and scale constraints to enforce consistent similarity relations across submaps. Experiments on TUM RGB-D and 7-Scenes show that UniSim-SLAM achieves state-of-the-art accuracy in the uncalibrated setting, reducing trajectory error by

38.5\%

on TUM RGB-D and

45.9\%

on 7-Scenes compared to prior best results. Project page: https://vision3d-lab.github.io/unisim-slam/

Abstract

Explore similar work

UniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) Optimization

ScaRF-SLAM: Scale-Consistent Reconstruction with Feed-Forward Models and Classical Visual SLAM

GeoGS-SLAM: Online Monocular Reconstruction Using Gaussian Splatting with Geometric Priors