AMB3R-SLAM: Kilometer-scale SLAM with Hierarchical Backend
Hengyi Wang, Lourdes Agapito
cs.CV, cs.RO
2026-09-17
AMB3R-SLAM tracks kilometer-scale monocular trajectories over 10k frames on one GPU. VBR ATE falls from 26.65 m to 7.42 m; LiDAR is sub-meter.
Geometric foundation models reconstruct well when they attend bidirectionally over a few dozen frames. Visual SLAM needs a stream, low latency, and memory that does not grow with trajectory length. Causal attention and recursive memory flatten the cost and give away local accuracy plus drift. AMB3R-VO already keeps full bidirectional attention inside a small active keyframe set, but it is still odometry: no global constraints, so kilometer routes drift.
AMB3R-SLAM hangs a hierarchical backend off that VO. Overlapping submaps fix the local window, sparse long-context mapping adds mid-range edges, loop closure pulls the trajectory together. Bundle adjustment that assumes a static world is skipped, so dynamic scenes run on the same pipeline. Stereo, RGB-D, and LiDAR plug in at the graph.
The frontend is DA3-Small, 80M parameters. Each new frame sees an anchor keyframe plus the two most recent frames; a robust solver recovers scale. Frontend confidence guides backend view selection, and backend poses re-anchor the frontend to kill local drift. On KITTI the in-window rotation error is 1.41 with 4.43% relative ATE at about 78 FPS.
The backend builds a dense submap of n frames every Δ = n/3 steps, so each submap overlaps both its neighbor and the one after that, which yields span-2 edges. Inside a submap, frames are grouped in time and only the most confident one per group is reconstructed; the rest are interpolated. A large foundation model runs on that compact set with the middle frame as reference. Every M submaps, a sparse window covering K submaps emits Sim(3) edges with |i−j| ≥ 3, dropped if confidence or covisibility is weak. Loops are proposed by DBoW2, verified by joint reconstruction plus voxel occupancy overlap, and rejected if the implied rotation or translation is implausible.
The pose graph optimizes keyframe Sim(3). Translation residuals are whitened by baseline length so long edges do not dominate in metres; Huber down-weights outliers. Stereo and RGB-D pin each submap to metric scale and shrink the graph to SE(3). LiDAR edges come from ICP, initialized at loops by the foundation model, with weights inverse-square in inlier RMSE.
Nine datasets, from tabletop metres to 5 km of city driving.
| Setting | This method | Prior online best |
| KITTI mono mean ATE | 13.11 m | LoGeR 18.65 m |
| VBR mono | 7.42 m (7.15 with Ω) | LingBot-Map 26.65 m, about −72% |
| Oxford Spires | 2.32 m (2.08 with Ω) | LoGeR 8.59 m, about −73% |
| TUM RGB-D | 2.3 cm | AMB3R-VO 3.2 cm |
| Bonn dynamic | 1.3 cm | WildGS-SLAM 2.3 cm |
| KITTI + LiDAR | 0.95 m | PIN-SLAM 1.22 m |
| VBR + LiDAR | 0.36 m | PIN-SLAM 0.81 m |
ETH3D AUC@5 cm moves from ViPE’s 38.7% to 44.0%. On fast EuRoC drones, AUC@5 cm is 20.3 versus offline VidMap at 11.5. All eight Bonn dynamic sequences rank first with no motion segmentation.
On an RTX 4090, monocular runs at 10.2–17.6 FPS and LiDAR at 31.3–47.8, peak memory 10–14 GB, independent of sequence length. Without loop closure, VBR full-trajectory window AUC falls from 90.39% to 42.02%. Dropping long-context mapping hurts the 50–1000 m windows; dropping whitening hurts every window. KITTI-01, 2.5 km of high-speed highway with poor texture, is still 55.91 m monocular and 2.95 m with LiDAR.
This is the paper that takes feed-forward geometry from “VO that runs” to “SLAM that survives kilometres,” with error cuts large enough to matter. For robotics, 10 FPS on one card with a memory cap is an installable budget. Dynamic scenes need no extra segmentation head. The LiDAR variant treats the foundation model as a loop initializer, not a replacement for ICP.
The map is a point cloud, so loops can ghost and duplicate surfaces; there is no compact surface. Extra sensors attach only at the graph, while the backbone still eats RGB. KITTI-01 still collapses in mono. The Ω backend wins on Oxford and VBR, so the default giant DA3 is not uniformly best. Real-time depends on window and sampling knobs that were not swept. There is no power comparison against a classical visual-inertial system with loop closure on the same kilometre routes.