PhysX-Omni: Unified Physical 3D Generation for Rigid, Deformable, and Articulated Objects

1 Introduction

PhysX-Omni is a new unified framework that generates simulation‑ready physical 3D assets for rigid, deformable and articulated objects. Unlike prior 3D AIGC methods that focus only on appearance, it models absolute scale, material, kinematics, affordance and textual description, enabling “interactive, movable, simulatable” assets for Physical AI and embodied AI research.

2 Method

2.1 Physical Geometry Representation

The authors introduce a template‑based Run‑Length Encoding (RLE) representation inspired by 2‑D RLE. Assets are voxelized and split into part‑level voxels; each part is sliced along the z‑axis into binary masks, which are encoded as compact 2‑D RLE tokens. Template layers allow multiple similar slices to share a template and store only residuals, dramatically reducing token count while preserving fine geometry. This explicit 3‑D representation avoids error accumulation common in autoregressive geometry generators.

2.2 PhysXVerse Dataset

To address the scarcity of simulation‑ready physical 3D data, the team built PhysXVerse, the first general‑purpose dataset containing over 8.7 K high‑quality assets spanning more than 2.9 K categories (indoor furniture, drones, robots, vehicles, large scene components, etc.). Assets were derived from PartVerse’s human‑validated part segmentation and further annotated with physical properties via a human‑in‑the‑loop pipeline.

2.3 PhysX‑Bench Benchmark

PhysX‑Bench is a unified benchmark that evaluates generated assets on six dimensions: Geometry, Absolute Scale, Material, Affordance, Kinematics and Description. Evaluation uses a vision‑language model (Qwen 3.5) and physics‑based simulation; metrics include CLIP alignment, multi‑view consistency, visual quality, and physical property assessments such as density and Young’s modulus via free‑fall and underwater‑drop videos.

3 Experiments

3.1 Results on Traditional Metrics

PhysX‑Omni was compared with PhysXGen, Articulate‑Anything, MonoArt and PhysX‑Anything on the PhysXVerse and PhysX‑Mobility datasets. It achieved the best scores on almost all geometry and physical‑attribute metrics. Notably, absolute‑scale error was reduced by two orders of magnitude relative to competing methods, and kinematics performance showed a marked improvement, indicating superior reasoning about joint structures and motion constraints.

3.2 Results on PhysX‑Bench

On the ground‑truth‑free PhysX‑Bench, PhysX‑Omni again led in most physical attributes, especially Absolute Scale, Material, Affordance, Kinematics and Description. Visualizations demonstrate higher robustness on complex structures, fine geometry and challenging articulated objects.

3.3 Applications

The generated assets were directly deployed in physics simulators for robot interaction and strategy learning, confirming their usefulness for large‑scale embodied‑AI data creation. The authors also explored scene‑level generation, showing potential for building whole‑scene simulation environments.

4 Conclusion

PhysX‑Omni delivers the first unified simulation‑ready physical 3D generation framework, producing assets enriched with geometry, material, kinematics and affordance information. Combined with the PhysXVerse dataset and PhysX‑Bench benchmark, extensive experiments demonstrate superior performance over existing methods and open new research directions for Physical AI, robotics and embodied AI.

Paper: https://arxiv.org/abs/2605.21572<br/>Code: https://github.com/physx-omni/PhysX-Omni