arXiv | Website | Video

This repository contains the implementation of ArtiPoint and the Arti4D dataset:

Articulated Object Estimation in the Wild

Abdelrhman Werby*, Martin Büchner*, Adrian Röfer*, Chenguang Huang, Wolfram Burgard and Abhinav Valada.
*Equal contribution.

Conference on Robot Learning (CoRL), 2025.

Follow these steps to install and run ArtiPoint.

1. Install environment

git clone https://github.com/robot-learning-freiburg/artipoint.git
cd artipoint
conda env create -f environment.yaml
conda activate artipoint
pip install -e .

2. Download model weights

mkdir -p checkpoints && cd checkpoints

# Mobile-SAM
gdown --fuzzy "https://drive.google.com/file/d/1dE-YAG-1mFCBmao2rHDp0n-PP4eH7SjE/view?usp=sharing"
unzip weight.zip

# CoTracker (offline and/or online)
wget https://huggingface.co/facebook/cotracker3/resolve/main/scaled_offline.pth
wget https://huggingface.co/facebook/cotracker3/resolve/main/scaled_online.pth || true

cd ..

3. Run on a dataset

# Minimal run (uses defaults)
artipoint

# Specify dataset root (Arti4D path)
artipoint dataset.root_path=/path/to/arti4d/raw

# Common options
artipoint \
   dataset.root_path=/path/to/arti4d/raw \
   debugging.write_results=true \
   visualization.show_3d_tracks=true

Please choose a desired top-level folder, download, and unzip the following scene-wise splits as well as the metadata:

wget http://aisdatasets.cs.uni-freiburg.de/arti4d/arti4d-din080.zip
wget http://aisdatasets.cs.uni-freiburg.de/arti4d/arti4d-rh078.zip
wget http://aisdatasets.cs.uni-freiburg.de/arti4d/arti4d-rh201.zip
wget http://aisdatasets.cs.uni-freiburg.de/arti4d/arti4d-rr080.zip
wget http://aisdatasets.cs.uni-freiburg.de/arti4d/arti4d-meta.zip

Dataset Details:

• The Arti4D splits are named din080, rh078, rh201, rr080. For each scene split you will find 8 to 16 sequences.
• The overall directory structure follows this format: arti4d/raw/SCENE/SEQUENCE/. Each scene defines a certain environment whereas the sequences are recordings within that particular scene.
• Difficulty levels and axis types are contained in arti4d/raw/metadata.yaml.
• Within each sequence folder you will find: - Depth and RGB data under depth / rgb. - The interaction intervals are defined in matched_cues.csv. - The GT camera odometry is provided under odom. - We also provide a mesh and point cloud reconstructions generated via TSDF-fusion: compressed_mesh.ply / compressed_point_cloud.ply - Furthermore, there is also a compiled video of the demonstration sequence contained.

At the moment we do only support the raw data download. Please contact buechner@cs.uni-freiburg.de in case you require the rosbag data.

Predicted camera odometry Furthermore, we also provide the camera odometry of DROID-SLAM and its registrations to the recorded maps using KISS-MATCHER:

wget http://aisdatasets.cs.uni-freiburg.de/arti4d/droid-slam-outputs.zip
wget http://aisdatasets.cs.uni-freiburg.de/arti4d/droid-slam-registrations.zip

If you find our work useful, please consider citing our paper:

@article{werby2025articulated,
   author={Werby, Abdelrhman and Buechner, Martin and Roefer, Adrian and Huang, Chenguang and Burgard, Wolfram and Valada, Abhinav},
   title={Articulated Object Estimation in the Wild},
   journal={Conference on Robot Learning (CoRL)},
   year={2025},
}

• Initial code release
• RIPL factor graph implementation
• Evaluation protocol