Converting static 3D meshes into interactable articulated assets is crucial for embodied AI and robotic simulation. However, existing zero-shot pipelines struggle with complex assets due to a critical lack of physical grounding. Specifically, ungrounded Vision-Language Models (VLMs) frequently suffer from kinematic hallucinations, while unconstrained joint estimation inevitably leads to catastrophic mesh inter-penetration during physical simulation. To bridge this gap, we propose MotionAnymesh, an automated zero-shot framework that seamlessly transforms unstructured static meshes into simulation-ready digital twins. Our method features a kinematic-aware part segmentation module that grounds VLM reasoning with explicit SP4D physical priors, effectively eradicating kinematic hallucinations. Furthermore, we introduce a geometry-physics joint estimation pipeline that combines robust type-aware initialization with physics-constrained trajectory optimization to rigorously guarantee collision-free articulation. Extensive experiments demonstrate that MotionAnymesh significantly outperforms state-of-the-art baselines in both geometric precision and dynamic physical executability, providing highly reliable assets for downstream applications.
Figures & tables
Figure 1: MotionAnymesh: Physics-Grounded Articulation of Static 3D Assets. While current SOTAs (e.g., Articulate-AnyMesh) rely on ungrounded semantics and often suffer from severe inter-penetration, MotionAnymesh ensures physically plausible articulation. Through kinematic-aware perception and physics-constrained optimization , our zero-shot framework transforms static meshes into collision-free, simulation-ready URDF digital twins for direct deployment in embodied AI tasks.
Figure 2: Overview of the MotionAnymesh framework. Our pipeline consists of three integrated stages: (1) Kinematic-Aware Part Segmentation, which extracts 3D-native primitives and clusters them using SP4D kinematic priors and VLM reasoning; (2) Joint Estimation and Optimization, featuring type-aware geometric initialization and physics-constrained trajectory refinement to ensure collision-free articulation ; and (3) Simulation-Ready Asset Finalization, providing simulation-ready URDF models with preserved textures.
Figure 3: Qualitative comparison of articulated object modeling. Compared with Articulate-Anything and Articulate-AnyMesh, MotionAnymesh produces cleaner geometric boundaries and more accurate kinematic structures, especially for complex mechanical objects like the robot arm and multi-part lamps.
Method
Part Segmentation
Joint Estimation
Simulation
mIoU ↑
Count Acc ↑
Type Err ↓
Axis Err ↓
Pivot Err ↓
Executability ↑
PARIS [ 25 ]
0.17
0.23
0.67
1.56
1.14
11%
URDFormer [ 3 ]
0.21
0.33
0.72
1.31
1.53
21%
SINGAPO [ 24 ]
0.52
0.66
0.24
0.73
0.57
43%
Articulate-Anything [ 20 ]
0.47
0.61
0.21
0.86
0.64
46%
Articulate-AnyMesh [ 35 ]
0.59
0.74
0.35
0.64
0.44
35%
Table 1: Quantitative comparison against state-of-the-art baselines across diverse 3D assets. We evaluate our method against others on three core dimensions: Part Segmentation, Joint Parameter Estimation, and Physical Executability. ↑ indicates higher is better, and ↓ indicates lower is better. Best results are highlighted in bold .
Clustering Strategy
mIoU ↑
Count Acc ↑
Pure VLM Semantics (w/o SP4D)
0.68
0.81
SP4D-Guided Clustering (Ours)
0.86
0.92
Table 2: Ablation on SP4D kinematic priors. Removing SP4D guidance ("w/o SP4D") causes the VLM to rely solely on semantic logic, leading to severe kinematic hallucinations and over-segmentation.
Joint Estimation Method
Axis Err ↓
Pivot Err ↓
Executability ↑
Kinematic Initialization (w/o Opt.)
0.23
0.27
65%
Physics-Constrained Opt. (Ours)
0.12
0.10
87%
Table 3: Impact of physics-constrained optimization. While geometric initialization provides a reasonable starting point, our non-linear optimization is essential to eliminate micro-misalignments and guarantee high physical executability in simulators.
Figure 4: Versatile articulation across diverse domains. MotionAnymesh successfully processes handcrafted assets and AI-generated surface meshes, transforming unstructured geometry into interactive digital twins without manual intervention.
Figure 5: Real-to-Sim-to-Real application. (Left) A static mesh is reconstructed from a single image. (Middle) MotionAnymesh generates a collision-free URDF for policy learning in simulation. (Right) The learned manipulation policy is deployed onto a physical robot, validating the sim-to-real fidelity of our estimated parameters.