We introduce VGGT-SLAM++, a complete visual SLAM system that leverages the geometry-rich outputs of the Visual Geometry Grounded Transformer (VGGT). The system comprises a visual odometry (front-end) fusing the VGGT feed-forward transformer and a Sim(3) solution, a Digital Elevation Map (DEM)-based graph construction module, and a back-end that jointly enable accurate large-scale mapping with bounded memory. While prior transformer-based SLAM pipelines such as VGGT-SLAM rely primarily on sparse loop closures or global Sim(3) manifold constraints - allowing short-horizon pose drift - VGGT-SLAM++ restores high-cadence local bundle adjustment (LBA) through a spatially corrective back-end. For each VGGT submap, we construct a dense planar-canonical DEM, partition it into patches, and compute their DINOv2 embeddings to integrate the submap into a covisibility graph. Spatial neighbors are retrieved using a Visual Place Recognition (VPR) module within the covisibility window, triggering frequent local optimization that stabilizes trajectories. Across standard SLAM benchmarks, VGGT-SLAM++ achieves state-of-the-art accuracy, substantially reducing short-term drift, accelerating graph convergence, and maintaining global consistency with compact DEM tiles and sublinear retrieval.
Figures & tables
Figure 2: (A) DEM-based scene representation on the TUM RGB-D freiburg1 teddy dataset. The DEMs provide a compact 2.5D encoding retaining geometric structure. (B) DEM visualizations from the 7-Scenes dataset. (C) DEMs generated for the Virtual KITTI (Sequence 01) dataset. (D) A full KITTI Odometry (Sequence 09) sequence demonstrating a complete loop, illustrating the ability of DEMs and our SLAM back-end pipeline to maintain global consistency.
Figure 3: Complete pipeline overview. The proposed VGGT-SLAM++ system comprises three main components: (a) Front-end: A Sim(3) odometry module that optimizes the relative poses of submaps generated by the feed-forward VGGT network. (b) Covisibility graph construction: A DEM-based map representation is used to compute structure-aware embeddings leveraging DINOv2, and an averaged tile score is used to insert spatially consistent nodes and edges into the covisibility graph. (c) Back-end: An optimization module that organizes submaps into a spatial graph and performs optimization over the detected spatial constraints.
Figure 4: Preservation of geometric cues in the DEMs, alongside semantics help in accurate trajectory estimation in planar scenes like the TUM RGB-D foor scene as shown in the figure.
Figure 5: (A), (B), (C): zero-shot object detection from DEMs proving structure preservation. (D) DEM of TUM-teddy and (E) color coding.
Figure 6: VGGT-SLAM++ results for : (A) custom data (406.8m) recorded by GoPro HERO10 camera with GPS groundtruth with 2m precision. (ATE RMSE 18±2 m); (B) custom data (1.8m) recorded by a OAK-1 camera with a Humanoid robot kinematics groundtruth (ATE RMSE 0.02m); (C) custom data (1.8m) recorded by a OAK-1 camera with Cobot forward kinematics groundtruth (ATE RMSE 0.01m); (D) KITTI Odometry 06 sequence (1230 m; ATE RMSE 13.65m); (E)TUM RGB-D 360 scene (5.82m; ATE RMSE 0.042m); (F) custom data (287.381m) path recorded by the GoPro HERO10 camera with GPS groundtruth with 2m precision. (ATE RMSE 7.17±2 m).
Method
Uncalib.
00
01
02
03
04
05
06
07
08
09
10
Avg.
Classical feature-based SLAM
ORB-SLAM2 (w/o LC) [ 55 ]
✗
40.65
502.20
47.82
0.94
1.30
29.95
40.82
16.04
43.09
38.77
5.42
69.73
ORB-SLAM2 (w/ LC) [ 55 ]
✗
6.03
508.34
14.76
1.02
1.57
4.04
11.16
2.19
38.85
8.39
6.63
54.82
LDSO [ 25 ]
✗
9.32
11.68
31.98
2.85
1.22
5.10
13.55
2.96
129.02
21.64
17.36
22.43
Learning-based SLAM
DROID-SLAM [ 77 ]
✗
92.10
5344.60
107.61
2.38
1.00
118.50
62.47
21.78
161.60
72.32
118.70
554.82
Table 1: KITTI Odometry Benchmark. Absolute trajectory RMSE error (ATE, meters). Gray shade indicates results from classical methods. “–” indicates SL(4) does not converge.
Method
Uncalib.
360
desk
desk2
floor
plant
room
rpy
teddy
xyz
Avg.
ORB-SLAM3 [ 9 ]
✗
0.017
0.210
–
0.034
–
–
–
0.009
N/A
N/A
DeepV2D [ 75 ]
✗
0.243
0.166
0.379
1.653
0.203
0.246
0.105
0.316
0.064
0.375
DeepFactors [ 14 ]
✗
0.159
0.170
0.253
0.169
0.305
0.364
0.043
0.601
0.035
0.233
DPV-SLAM [ 44 ]
✗
0.112
0.018
0.029
0.057
0.021
0.330
0.030
0.084
0.010
0.076
DPV-SLAM++ [ 44 ]
✗
0.132
0.018
0.029
0.050
0.022
0.096
0.032
0.098
0.010
0.054
GO-SLAM [ 94 ]
✗
0.089
0.016
0.028
0.025
0.026
0.052
0.019
0.048
0.010
0.035
Table 2: TUM RGB-D Benchmark. Absolute trajectory RMSE error (ATE, meters). Gray shade indicates results from calibrated methods.
Table 4: Virtual KITTI Benchmark. RMSE ATE (m). Methods shown per sequence, across weather variants.
Method
00
01
02
03
04
05
06
07
08
09
10
Avg.
VGGT-SLAM++ (default)
119.00
109.64
223.21
4.50
0.95
25.21
13.65
12.17
155.00
35.26
15.71
64.936
Mean reducer (no softmax)
120.49
109.64
223.21
4.50
0.95
25.21
13.65
12.17
155.00
35.26
15.71
65.072
Softmax ( τ=0.10 )
119.00
109.64
223.21
4.50
0.95
25.21
13.65
12.17
155.00
35.26
15.71
64.936
Softmax ( τ=0.005 )
120.49
109.64
219.25
4.50
0.95
25.21
13.65
12.17
155.00
35.26
15.71
64.711
Half resolution
116.53
109.64
171.42
4.50
0.95
20.79
5.85
12.17
155.00
35.26
15.71
58.893
High resolution
132.54
109.64
223.21
4.50
0.95
23.32
13.65
12.17
155.00
35.26
15.71
65.995
Table 5: DEM hyperparameter ablations on KITTI Odometry.
Method
Uncalib.
MH01
MH02
MH03
MH04
MH05
Avg.
ORB-SLAM [ 54 ]
✗
0.071
0.067
0.071
0.082
0.060
0.070
DSM [ 98 ]
✗
0.039
0.036
0.055
0.057
0.067
0.051
ORB-SLAM3 [ 54 ]
✗
0.016
0.027
0.028
0.138
0.072
0.056
DeepFactors [ 14 ]
✓
1.587
1.479
3.139
5.331
4.002
3.108
DROID-SLAM [ 77 ]
✓
0.013
0.014
0.022
0.043
0.043
0.027
VGGT-SLAM (Sim(3)) [ 50 ]
✓
1.740
2.890
2.270
3.390
4.400
2.938
Table 6: EuRoC MAV Benchmark. Absolute trajectory RMSE error (ATE, meters). Gray shade indicates results from calibrated methods. “–” indicates SL(4) does not converge.
Figure 7: The red line is the ground truth reference from GPS readings and the blue line is the estimation by VGGT-SLAM++ for the custom GoPro camera dataset [Axes’ units are in meters].
Figure 8: The DEM rendered from the 3D points aligned by odometry over the KITTI Sequence 05, with color mapping for better visualisation.
HFIPS, Chinese Academy of Sciences, Hefei, China. · University of Science and Technology of China, Hefei, China. · Aarhus University, Aarhus, Denmark. +3