Artic-O

End-to-End Articulated Object Reconstruction via Latent Geometry Learning

Xuyang Wang^1,2,* Zhenyu Li² Jian Ding^2,† Habib Slim² Peter Wonka² Hongdong Li¹ Mohamed Elhoseiny²

¹Australian National University ²KAUST

* Work done during the author's research internship at KAUST. † Project lead.

SIGGRAPH Asia 2026 · Kuala Lumpur · December 1–4, 2026

Artic-O takes sparse images of an object at an open and a closed state and predicts complete geometry, the movable part, and joint parameters in 0.32 seconds, then poses the object at arbitrary articulation states. — From sparse images at two articulation states, **Artic-O** predicts complete geometry, the active part, and joint parameters in a single feed-forward pass (**0.32 s**). Forward kinematics then poses the object at any intermediate state. Right: quality against inference cost versus prior work.

Abstract

Reconstructing articulated objects from sparse images requires recovering complete geometry, movable parts, and motion parameters. Recent methods typically separate geometry reconstruction, part reasoning, and articulation estimation into different stages. This separation can weaken consistency between shape, active parts, and motion, while also incurring substantial inference cost.

We introduce Artic-O, an end-to-end, feed-forward framework for articulated object reconstruction via latent geometry learning. Instead of fitting geometry in image or view space, Artic-O maps sparse multi-state observations into a pretrained latent geometry space, where a frozen flow-matching decoder provides a complete-shape prior for recovering visible and occluded structures. To connect geometry with articulation, Artic-O fuses visual tokens, geometry latents, and point-wise decoder features in an image-grounded part-reasoning module for active-part segmentation and articulation prediction. We further train the model with a geometry-to-articulation curriculum and a decoupled two-pass strategy to balance reconstruction and part-level supervision.

On PartNet-Mobility, Artic-O achieves strong reconstruction quality while being substantially more efficient than LARM, a strong prior method. It reduces Chamfer Distance, improves F-score, and achieves comparable or better articulation accuracy across most joint metrics, while reducing inference time from 9 minutes to about 0.3 seconds per object.

Method

Artic-O architecture: a state-aware image encoder built on DINOv2 with state embedding injection, a geometry latent attended jointly with image tokens, a frozen flow-matching shape decoder, and an image-grounded segmentation module feeding articulation heads. — A **state-aware image encoder** injects a learned state embedding into DINOv2 features and attends them jointly with a geometry latent. A **frozen shape decoder** turns that latent into a complete point cloud, supplying the completion prior. An **image-grounded part-reasoning module** — an image memory bank plus PAT slot blocks — splits context from active slots, driving the segmentation and articulation heads.

Results

0.32 s

per object, versus 9 min for LARM

−15%

Chamfer Distance versus LARM

+2.4 pp

F1@0.05 versus LARM

Metric	Artic-O	LARM
Chamfer Distance ↓	0.01661	0.0196
F1@0.05 ↑	0.9578	0.9337
Axis angle @0.25 ↑	0.9490	0.9570
Axis origin @0.15 ↑	0.9804	0.9690
Motion range @0.3 ↑	0.9216	0.8550

PartNet-Mobility, averaged over five articulation states, on the LARM-comparable subset (n = 255) with deterministic view selection. Artic-O leads on four of five reported metrics.

Qualitative comparison against LARM and ground truth at articulation states 0.25 and 0.75 for a cabinet, a shelf, a trash can, and a drawer. — Qualitative comparison at states 0.25 and 0.75. Context geometry in blue, the predicted movable part in amber, the estimated joint axis in red (revolute) or green (prismatic). Artic-O recovers interior structure — shelves, cavities, drawer boxes — that view-conditioned reconstruction leaves hollow or broken.

Interior geometry

Two objects, a shelf unit and a chest of drawers, each shown with a cut plane through the reconstruction and the resulting cutaway. LARM, Artic-O and ground truth are compared side by side; LARM's interiors are largely empty while Artic-O reproduces the shelves and dividers present in the ground truth. — Slicing the reconstructions open. The top row of each pair places a cut plane through the object; the bottom row shows the geometry behind it. LARM leaves the interior hollow — the shelves and dividers simply are not there, because nothing in the input views observes them. Artic-O completes them, closely matching the ground-truth interior.

In the wild

Two real objects photographed by hand — a kitchen wall cabinet and a bedside table — each shown at two articulation states, alongside the predicted articulated reconstruction at states 0.25 and 0.75 with the recovered hinge axis drawn in red. — Handheld photographs of a kitchen wall cabinet and a bedside table, with no calibration, no camera poses and no ground truth. From three casual shots at each of two states, Artic-O separates the door from the body, recovers the interior shelving, and places a plausible hinge axis (red). The reconstructions are then posed at intermediate states the photographs never show.

Code and models

The release includes evaluation and inference code, the trained checkpoint, and the test split — enough to reproduce the benchmark table above and to run the model on a test object end to end. Training code and the data-construction pipeline are not included.

git clone https://github.com/Wxyxixixi/Artic-O.git
cd Artic-O
pip install -r requirements.txt
bash scripts/fetch_third_party.sh

hf download wxyxixixi/artic-o artic_o_s0_36.pth --local-dir ./checkpoints/
hf download wxyxixixi/artic-o-data --repo-type dataset --local-dir ./data/

bash scripts/verify_benchmark.sh

Citation

@article{wang2026artic,
  title={Artic-O: End-to-End Articulated Object Reconstruction via Latent Geometry Learning},
  author={Wang, Xuyang and Li, Zhenyu and Ding, Jian and Slim, Habib and Wonka, Peter and Li, Hongdong and Elhoseiny, Mohamed},
  journal={arXiv preprint arXiv:2606.21938},
  year={2026}
}