Project Resources

Paper PDF

Open the arXiv PDF for MotionAnymesh.

arXiv
Main LaTeX Source

Included from the uploaded ECCV2026 package.

Source

MotionAnymesh: Physics-Grounded Articulation for Simulation-Ready Digital Twins

Zero-shot conversion of static 3D meshes into collision-free, simulation-ready articulated URDF digital twins.

WenBo Xu1, Liu Liu1, Li Zhang2, Dan Guo1, RuoNan Liu3
1Hefei University of Technology 2The Hong Kong Polytechnic University 3Shanghai Jiao Tong University
ECCV 2026 Submission
Articulated Object Modeling Kinematic Segmentation Simulation-Ready URDF
Paper Code (coming soon)
MotionAnymesh teaser showing static meshes converted into articulated digital twins
MotionAnymesh transforms static 3D assets into collision-free, simulation-ready URDF digital twins through kinematic-aware perception and physics-constrained optimization.

Abstract

Converting static 3D meshes into interactable articulated assets is crucial for embodied AI and robotic simulation. Existing zero-shot pipelines struggle with complex assets because ungrounded VLM reasoning can hallucinate kinematic structures, while unconstrained joint estimation often causes severe inter-penetration during simulation.

MotionAnymesh is an automated zero-shot framework that transforms unstructured static meshes into simulation-ready digital twins. It combines kinematic-aware part segmentation grounded by SP4D physical priors with geometry-physics joint estimation, including type-aware initialization and physics-constrained trajectory optimization. Extensive experiments show stronger geometric precision and dynamic physical executability than prior baselines.

Highlights

MotionAnymesh addresses the missing physical grounding in existing static-to-articulated pipelines by unifying kinematic-aware perception with physics-constrained actuation.

1

Kinematics-Grounded 3D Part Segmentation

We decouple physical boundary recovery from semantic grouping: clean 3D-native primitives preserve object geometry, while explicit multi-view kinematic priors guide their assembly into coherent movable parts, suppressing both 2D projection artifacts and VLM kinematic hallucinations.

2

Geometry-Physics Joint Optimization

We formulate joint estimation as a physically constrained optimization problem: type-aware contact-interface initialization predicts robust axes and pivots, then trajectory refinement penalizes collision, detachment, and micro-misalignment throughout virtual articulation.

3

Simulation-Ready Digital Twins

We convert the recovered part hierarchy, optimized joints, motion limits, and preserved textures into standard URDF assets that can be directly executed in physics simulators and used for embodied AI and real-to-sim-to-real robotic workflows.

Method Overview

MotionAnymesh pipeline overview
The framework consists of kinematic-aware part segmentation, joint estimation and optimization, and simulation-ready asset finalization with texture preservation.

Qualitative Comparison

Qualitative comparison with articulation baselines
Compared with Articulate-Anything and Articulate-AnyMesh, MotionAnymesh produces cleaner boundaries and more accurate kinematic structures for complex objects.

Applications

MotionAnymesh applications across diverse 3D asset domains
MotionAnymesh processes handcrafted assets and AI-generated surface meshes, turning unstructured geometry into interactive digital twins.
Real-to-sim-to-real robotic application
A real-to-sim-to-real workflow reconstructs a static mesh from a real image, builds a collision-free URDF, trains a policy in simulation, and deploys it on a physical robot.

Qulitative Results

Qualitative results on PartNet-Mobility
Extended qualitative results on PartNet-Mobility objects.
Qualitative results on Objaverse objects
Extended qualitative results on Objaverse objects.
Qualitative results on generative 3D assets
Extended qualitative results on generative 3D assets.

BibTeX

@misc{xu2026motionanymeshphysicsgroundedarticulationsimulationready,
      title={MotionAnymesh: Physics-Grounded Articulation for Simulation-Ready Digital Twins}, 
      author={WenBo Xu and Liu Liu and Li Zhang and Dan Guo and RuoNan Liu},
      year={2026},
      eprint={2603.12936},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2603.12936}, 
}