AMB3R-SLAM: Kilometer-scale SLAM
with Hierarchical Backend

@article{wang2026amb3rslam,
  title={AMB3R-SLAM: Kilometer-scale SLAM with Hierarchical Backend},
  author={Wang, Hengyi and Agapito, Lourdes},
  journal={arXiv preprint arXiv:2609.19518},
  year={2026}
}

1. Clone the repository

git clone https://github.com/HengyiWang/amb3r-slam.git
cd amb3r-slam

2. Create a conda environment

conda create -n amb3r-slam python=3.9 -y
conda activate amb3r-slam
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118   # match your CUDA
pip install -r requirements.txt
bash thirdparty/build.sh  

3. Download weights

bash checkpoints/download.sh       # WITH_SALAD=1 for the SALAD checkpoint

This fetches DA3NESTED-GIANT-LARGE-1.1, DA3-SMALL and ORBvoc.txt into ./checkpoints/ (plus dino_salad.ckpt with WITH_SALAD=1). --model_name omega also needs checkpoints/vggt_omega_1b_512.pt.

python run.py --dataset demo --data_path <path-to-image-folder> --save_map

Results are written to outputs/demo_mono_<model>/:

• traj_<folder>.npz : the trajectory, one camera-to-world pose per frame
• map_<folder>.ply : a coloured point cloud of the map, every pixel (with --save_map)
• map_<folder>_stride.ply : the same map at one pixel in --pixel_stride along each axis (with --save_map)

• --pixel_stride n : the pixel stride of the _stride cloud (default 4)
• --save_every n : keep one of every n mapped frames (default 2)
• --map_max_points n : cap the size of the cloud
• --fps r : the capture rate, if the frames were captured below 10 Hz, please specify it to avoid submap geometry issues (default 10 Hz)
• --stride n : keep every n-th frame

We keep the command used for each dataset:

bash scripts/run_eval.sh [DATASET ...]

You can run each individual dataset like:

# monocular
python run.py --dataset tum   --data_path $TUM
python run.py --dataset kitti --data_path $KITTI

# RGB-D / stereo
python run.py --dataset tum   --data_path $TUM   --sensor rgbd
python run.py --dataset kitti --data_path $KITTI --sensor stereo

# LiDAR
python run.py --dataset kitti --data_path $KITTI --sensor lidar
python run.py --dataset vbr   --data_path $VBR   --sensor lidar

• --sensor rgbd|stereo|lidar : the input modality (default monocular)
• --model_name omega : use VGGT-Ω instead of Depth Anything 3
• --calib : give the dataset's intrinsics to the system. It runs uncalibrated otherwise

The AMB3R-SLAM pipeline (everything outside thirdparty/) is released under the Apache License 2.0. The third-party packages under thirdparty/ and the model weights are not covered by it; please follow the license of each one respectively.

Our code builds upon several amazing open-source projects:

• Models: AMB3R, Depth Anything 3, VGGT-Ω
• Loop retrieval & matching: DBoW2, DPV-SLAM, SALAD, MegaLoc, ALIKED
• LiDAR & evaluation: KISS-ICP, evo

We sincerely thank the authors for their open-source contributions!

If you find our code or paper useful, please consider citing:

@article{wang2026amb3rslam,
  title={AMB3R-SLAM: Kilometer-scale SLAM with Hierarchical Backend},
  author={Wang, Hengyi and Agapito, Lourdes},
  journal={arXiv preprint arXiv:2609.19518},
  year={2026}
}