DailyArt: Discovering Articulation from Single Static Images via Latent Dynamics

Hang Zhang; Qijian Tian; Jingyu Gong; Daoguo Dong; Xuhong Wang; Yuan Xie; Xin Tan

cs.CVApr 9, 2026

DailyArt: Discovering Articulation from Single Static Images via Latent Dynamics

Authors: Hang Zhang, Qijian Tian, Jingyu Gong, Daoguo Dong, Xuhong Wang, Yuan Xie, Xin Tan

Organizations: East China Normal University · Shanghai AI Lab · Shanghai Jiao Tong University · Fudan University

Abstract

Articulated objects are essential for embodied AI and world models, yet inferring their kinematics from a single closed-state image remains challenging because crucial motion cues are often occluded. Existing methods either require multi-state observations or rely on explicit part priors, retrieval, or other auxiliary inputs that partially expose the structure to be inferred. In this work, we present DailyArt, which formulates articulated joint estimation from a single static image as a synthesis-mediated reasoning problem. Instead of directly regressing joints from a heavily occluded observation, DailyArt first synthesizes a maximally articulated opened state under the same camera view to expose articulation cues, and then estimates the full set of joint parameters from the discrepancy between the observed and synthesized states. Using a set-prediction formulation, DailyArt recovers all joints simultaneously without requiring object-specific templates, multi-view inputs, or explicit part annotations at test time. Taking estimated joints as conditions, the framework further supports part-level novel state synthesis as a downstream capability. Extensive experiments show that DailyArt achieves strong performance in articulated joint estimation and supports part-level novel state synthesis conditioned on joints. Project page is available at https://rangooo123.github.io/DaliyArt.github.io/.

Figures & tables

Fig. 1: Overview of DailyArt. We propose a synthesis-mediated framework for articulated joint parameter estimation and controllable motion synthesis from a single static image. Given an input image, DailyArt first synthesizes a maximally articulated (opened) state to reveal hidden kinematic cues, which helps reduce 2D ambiguity. DailyArt (1) estimates joint parameters (type, axis, and motion range) from cross-state discrepancies, and (2) enables part-level articulated state synthesis.

Fig. 2: Comparison of current pipelines (left & mid) and our proposed DailyArt (right) . Up: In joint estimation, existing pipelines leverage priors to guide the single image, multi-view or multi-state images. DailyArt generates novel opened-state images by encoding a state index into the image feature. The kinematic motion difference within dual-state images are directly compared and used to estimate the joints information. Down: For showing novel states of an object, some methods import URDF files into simulators. DailyArt uses both the input image and estimated joints to synthesis multiple states of every kinematic joint.

Fig. 3: Kinematic states of an articulated object . We define $t=0$ as the closed state where the part is closed or remains inactivated, and define $t=1$ as the opened state where the part reaches the maximum articulated limit. And the motion index $t=t^{\prime}$ is a condition to describe novel states of parts somewhere within the motion range.

Fig. 4: Overview of the DailyArt Framework. Given a single closed-state image $\mathbf{I}_{0}$ , DailyArt adopts a three-stage paradigm to estimate joints and synthesis images. The input is processed by a prior-free novel state synthesis (Stage I) to an opened-state $\hat{\mathbf{I}}_{1}$ , revealing occluded motion evidence. For joint estimation (Stage II), the input and synthesized states are lifted into 3D point-maps as $(\mathcal{P}_{0},\mathcal{P}_{1})$ to estimate a set of joint parameters $\hat{\mathcal{J}}$ . Having estimated joints, DailyArt could extend the novel state synthesis as joint-conditioned synthesis (Stage III) from the input $\mathbf{I}_{0}$ to novel states related to different joints $\hat{\mathbf{I}}_{t=t^{\prime},\mathcal{J}}$ . DailyArt performs outputs without part annotations or priors.

Method	Single Image	Extra Priors	Interaction	Multi-State
Novel State Synthesis Baselines
DragAPart	✓	–	Drag Points	–
PartRM	–	Drag from Multi-state Masks	–	Zero123+
Puppet-Master	✓	–	Drag Points	–
LARM	–	Multi-views	Camera Position	As Inputs
Joint Estimation Baselines

TABLE I: Task Level Comparisons. We disclose the input modalities and extra requirements for each baseline. DailyArt is the only method that enables both high-fidelity synthesis and precise kinematic estimation from a single static image without requiring interaction, retrieval, or language prompts.

Fig. 5: Unseen object test results. We test DailyArt performance on unseen objects. The images are segmented with a transparent background as inputs. The results demonstrate that DailyArt can handle such inputs and synthesise novel states.

Method	Overall $\uparrow$	Type $\downarrow$	Origin $\downarrow$	Direct. $\downarrow$	Range $\downarrow$
URDFormer [ 19 ]	48.6	0.342	0.188	0.370	0.335
Singapo [ 20 ]	35.4	0.482	0.285	0.512	0.440
Articulate-Anything [ 18 ]	56.0	0.288	0.165	0.322	0.310
Physx-Anything [ 41 ]	62.8	0.295	0.130	0.325	0.282
DailyArt (Ours)	68.4	0.215	0.124	0.275	0.242

TABLE II: PartNet-Mobility joint estimation. We report the Overall Success Rate (%,

\uparrow

) and mean errors (

\downarrow

) for individual joint attributes.

Method	Overall $\uparrow$	Type $\downarrow$	Origin $\downarrow$	Direct. $\downarrow$	Range $\downarrow$
URDFormer [ 19 ]	37.5	0.738	0.395	0.625	0.482
Singapo [ 20 ]	32.4	0.819	0.372	0.584	0.466
Articulate-Anything [ 18 ]	48.3	0.370	0.268	0.351	0.403
Physx-Anything [ 41 ]	52.8	0.338	0.204	0.377	0.371
DailyArt (Ours)	54.4	0.275	0.204	0.349	0.368

TABLE III: AKB-48 joint estimation. We report the zero-shot Overall Success Rate (%,

\uparrow

) and mean errors (

\downarrow

) for individual joint attributes.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIP-T $\uparrow$	FVD $\downarrow$
DragAPart [ 50 ]	21.2	0.837	0.143	0.632	212.4
PartRM [ 24 ]	22.8	0.840	0.145	0.643	219.5
Puppet-Master [ 26 ]	23.8	0.829	0.110	0.678	204.3
LARM [ 23 ]	24.3	0.907	0.104	0.749	205.4
DailyArt (Ours)	25.5	0.920	0.102	0.766	202.2

TABLE IV: PartNet-Mobility novel-state synthesis. We report the visual fidelity and semantic consistency of the synthesized opened state

\hat{\mathbf{I}}_{1}

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIP-T $\uparrow$	FVD $\downarrow$
DragAPart [ 50 ]	16.3	0.724	0.355	0.512	312.4
PartRM [ 24 ]	18.1	0.752	0.285	0.523	268.5
Puppet-Master [ 26 ]	17.8	0.815	0.249	0.534	246.1
LARM [ 23 ]	18.3	0.813	0.174	0.654	265.4
DailyArt (Ours)	19.6	0.821	0.162	0.656	245.2

TABLE V: AKB-48 novel-state synthesis. We report the zero-shot quality of the synthesized opened state

\hat{\mathbf{I}}_{1}

on real-world objects.

Fig. 6: Qualitative ablation results. Left: Without Stage 1 target-state synthesis, direct joint regression from a single closed-state image often fails to synthesise a plausible articulation state (the opened notebook is more like a laptop). Right: Without 3D lifting, a 2D pair encoder may appear reasonable in the image plane but produces incorrect joint geometry in 3D (inaccurate joint estimations could affect the kinematic part’s motion). These examples highlight the importance of both synthesis-mediated cross-state reasoning and 3D geometric constraints in DailyArt.

	Configuration	PSNR $\uparrow$	Overall SR $\uparrow$	Latency $\downarrow$
	DailyArt (Full Pipeline)	25.5	68.4%	0.45s
I. Pipeline
A	Direct Regression (No Synthesis)	–	44.2%	0.18s
A	Change vs. Full	–	-24.2%	-0.27s
B	Oracle Synthesis (GT Target State)	–	69.7%	0.20s
B	Gap to Upper Bound	–	+1.3%	–

TABLE VI: Ablation Study. We first validate the core two-stage pipeline design (I), and then study several design choices in the full model (II). The Full model provides the best overall balance among the non-oracle configurations evaluated here.

Fig. 7: Failure cases on real-world unseen object and part segmentation on articulated objects. Even when conditioned with optimal text or point prompts, off-the-shelf foundation segmentation models fail to separate moving parts from the static base. This suggests that priors and prompts are less effective for identifying kinematic structures than we thought.

Fig. 8: Visual Comparison on Joint-conditioned Novel State Synthesis (Stage III) of DailyArt and baselines. We prepared priors for baselines, such as drags (calculated from input and gt meshs), seg masks from LLM, and camera extrinsic.

Fig. 9: Visual Comparison on Joint Estimation (Stage II). The visualization results differ due to the variations on how methods predict the joint parameters, including engine annotations, mesh building files or 3D coordinates. Instead of generating only a URDF structure for the simulation engine or part retrievals, DailyArt estimates joints based on the current view, and provides part control information possible for both generative pipelines or robot interactions, including motion ranges and axis directions.

Articulated Object Reconstruction

Articulated object reconstruction recovers 3D part geometry, joints, and motion from images, video, or depth for digital twins and robot interaction.

See all 60 papers

Query-Conditioned Articulation Estimation from a Single Image
Oct 1, 2026Abdelrhman Werby, Fabio Scaparro, Kai O. ArrasArticulated Object Reconstruction Robotic Manipulation
Socially Intelligent Robotics Lab, Institute for Artificial Intelligence, University of Stuttgart, Germany
PAct: Part-Decomposed Single-View Articulated Object Generation
Feb 16, 2026Qingming Liu, Xinyue Yao, Shuyuan Zhang +4Articulated 3D Asset Generation 3D Asset Generation
The Chinese University of Hong Kong, Shenzhen, China · DexForce Technology, China
Instruct-Particulate: Scaling Feed-Forward 3D Object Articulation with Kinematic Control
Jun 12, 2026Ruining Li, Yuxin Yao, Matt Zhou +5Articulated Object Reconstruction 3D Part Segmentation
University of Oxford · University of Cambridge · 3Nanyang Technological University

Oct 1, 2026J/K move · Enter open · S save

Query-Conditioned Articulation Estimation from a Single Image

Abdelrhman Werby, Fabio Scaparro, Kai O. Arras

Socially Intelligent Robotics Lab, Institute for Artificial Intelligence, University of Stuttgart, Germany

Articulated Object Reconstruction Robotic Manipulation Robotic Perception

Open paper AskPDF arXiv

Enabling robots to estimate the kinematic parameters of articulated objects unlocks a wide range of capabilities for interaction and manipulation. The estimation has to happen from the information the robot currently observes, often just a single RGB image of an object it has never seen before. Current single-image approaches couple articulation part segmentation with articulation estimation, making their predictions vulnerable to missed detections and incorrect part associations, and they regress metric 3D geometry that a single view fixes only up to scale. We present QueryArt, a model that estimates articulation parameters from a single RGB image, a 2D query point, and camera intrinsics. QueryArt is trained to estimate the 3D articulation geometry relative to the queried point and in units of its depth, which keeps its target identifiable from the image alone. A single depth measurement at the query point then supplies the scale and recovers the metric parameters. We train QueryArt on a curated mixture of synthetic and real-world articulation datasets. We evaluate QueryArt on several benchmarks and compare it against recent baselines. QueryArt outperforms recent baselines on most articulation metrics, including on out-of-distribution data. To demonstrate the model's capabilities in real-world settings, we evaluate QueryArt on a mobile manipulator across 57 manipulation trials spanning 16 object parts and five viewpoint classes, achieving a 70.2% success rate. We provide code and videos at: https://abwerby.github.io/queryart/

Abstract

Articulated Object Reconstruction

Explore similar work

Query-Conditioned Articulation Estimation from a Single Image

PAct: Part-Decomposed Single-View Articulated Object Generation

Instruct-Particulate: Scaling Feed-Forward 3D Object Articulation with Kinematic Control