PhysX-Mobility: Articulated 3D Models Dataset

• PhysX-Mobility is a comprehensive dataset of articulated, simulation-ready 3D models annotated with explicit geometric, kinematic, and physical properties.
• It employs voxelization, surface mesh reconstruction, and run-length encoding to achieve efficient integration with popular physics engines.
• The dataset supports advanced research in embodied AI, robotic manipulation, and physics-based simulation through rigorous quality assurance and diverse asset representation.

PhysX-Mobility is a large-scale, publicly available dataset comprising articulated, simulation-ready 3D models of common real-world objects, each annotated with explicit geometric, kinematic, and physical properties. Introduced in “PhysX-Anything: Simulation-Ready Physical 3D Assets from Single Image” (Cao et al., 17 Nov 2025), PhysX-Mobility significantly extends the diversity and depth of previous datasets—such as PartNet-Mobility—thereby providing the first comprehensive repository suitable for direct use in physics engines for embodied AI research, robotic learning, and physics-based simulation.

1. Dataset Scope and Content

PhysX-Mobility consists of approximately 2,100 assets distributed over 47 object categories, more than doubling the category coverage of prior physically annotated 3D collections (Cao et al., 17 Nov 2025). These categories include a representative selection of household and office objects such as cabinet, coffee_machine, faucet, camera, stapler, fan, toilet, eyeglasses, lighter, and others. Each object is modeled with part-level articulation and detailed physical properties enabling realistic simulation of interaction and manipulation.

The asset distribution ranges from categories with as few as 20 unique models (for relatively rare objects) up to 80 instances in common classes. On average, each category contains approximately 45 assets. The objects' physical properties span a wide mass distribution: from lightweight items ($\sim10^{-3}$ kg for eyeglasses) up to 20 kg for heavy appliances (e.g., coffee machines), with a log-normal distribution (σ ≈ 1.2). Material diversity covers 12 canonical types, with an entropy of approximately 2.8 bits.

2. Annotation Schema and Data Structure

Each asset in PhysX-Mobility includes multi-level annotation to support direct import and reliable simulation across widely-used physics engines such as MuJoCo, Bullet, and Drake. Assets are structured as follows:

• Geometry:
  • Coarse voxels: A $32 \times 32 \times 32$ occupancy grid, with occupied voxel indices serialized and compressed via run-length encoding; this yields a 193× reduction in text token usage over naive mesh flattening.
  • Fine meshes: Part-level surface meshes (OBJ format) are reconstructed using latent diffusion decoders and assigned to their respective parts via nearest-neighbor mapping from voxels.
• Articulation:
  • Each movable part is annotated with its own joint record, including type (revolute, prismatic, fixed), axis (unit vector in object frame), pivot origin, and travel limits (in radians for revolute, meters for prismatic). All joints are single-DoF by construction.
• Physical Properties:
  • Mass ($m$), center of mass ($c$), density ($\rho$), inertia tensor ($I$ for principal moments), frictional coefficients (static $\mu_s$ and dynamic $\mu_d$), and restitution ($e$) are provided.
  • Densities are assigned from canonical material databases; mass and inertial properties are computed from the mesh, e.g., for principal inertia,

  $I_{ii} = \int_V (\|x\|^2 - x_i^2) \rho \, dV$

• Metadata and Formats:

  • Each object directory contains: part meshes, voxel serialization, URDF and XML files for simulators, and metadata.json capturing all annotations.

A representative JSON schema for metadata.json includes fields for id, category, part identifiers, geometry references, articulation array, physical parameter block (as above), and natural language description.

3. Construction Pipeline and Quality Assurance

PhysX-Mobility was created by augmenting the articulated model collection in PartNet-Mobility. The pipeline consists of:

1. Voxelization of geometry to $32^3$ occupancy grids and run-length serialization for efficient text and VLM integration.
2. Surface mesh reconstruction via structured generative models and part-wise segmentation.
3. Canonical material assignment by human annotators to support plausible density, friction, and restitution parameterization.
4. Analytic computation of mass, center of mass, and inertias from meshes.
5. Manual spot-checking of >10% of assets to verify consistency of joint axes, limits, and physical realism of annotations.

This process yields high-fidelity, simulation-ready assets. Material properties are drawn from a set of approximately 12 standard types, and parameters such as friction and joint limits are sampled or selected within realistic, category-specific ranges.

4. Statistical Properties

PhysX-Mobility’s statistical properties underscore its diversity and realism:

Statistic	Value/Range	Context
Categories	47	E.g., faucet, camera, cabinet, etc.
Model count	≈2,100	20–80 per category, avg. ≈45
Mass	$10^{-3}$–$20$ kg	Log-normal, σ ≈ 1.2
Inertia $I_x, I_y, I_z$	$10^{-6}$–$10^{-2}$ kg·m²
Friction $\mu_s$, $\mu_d$	[0.2, 0.6], [0.1, 0.5]
Joint range (revolute/prism.)	±$\pi/2$, $\pm0.15$ m	Mean spans
Material entropy	≈2.8 bits (over 12 types)

All models provide fine-grained per-part segmentation, explicit geometry and articulation, and dense, physically consistent parameterization for simulation use.

5. Dataset Format, Access, and Licensing

The dataset is structured for direct use and ease of integration into embodied AI pipelines. The directory hierarchy is organized by split (train/val/test), then by category and model_id:

PhysX-Mobility/ ├─ train/ │ └─ category_name/ │ └─ model_id/ │ ├ mesh/ # .obj + .mtl part meshes │ ├ voxels.txt # voxel occupancy serialization │ ├ model.urdf # articulated model for simulators │ ├ model.xml # simulator port │ └ metadata.json # full schema as above ├─ val/ … └─ test/ …

All assets, metadata, and importable simulator files are distributed at https://physx-anything.github.io under the CC-BY-4.0 license, supporting both open research and commercial applications (Cao et al., 17 Nov 2025). The test split is drawn from the full set; qualitative documentation makes it likely that ≈300 examples are reserved for held-out testing, with the remainder split into training and validation.

6. Use Cases, Evaluation Benchmarks, and Applications

PhysX-Mobility enables several core research activities:

• Generative evaluation: The PhysX-Anything framework was benchmarked using PhysX-Mobility assets, with reported geometry PSNR of 20.35, Chamfer Distance 14.43, F-score 77.5, absolute scale error 0.30 m, and kinematic parameter score 0.83. All metrics are normalized to facilitate comparison (Cao et al., 17 Nov 2025).
• In-the-wild generalization: Models trained on PhysX-Mobility evaluated on 100 Internet images attained VLM geometry and kinematics scores of 0.94. Human raters (n = 14, 1,568 ratings) preferred output geometries 98% of the time and all physical attributes over 94%.
• Simulation-based embodied learning: In MuJoCo-style simulation with the “robopal” 0.3.1 framework, learned policies for tasks such as faucet turning, cabinet opening, delicate pick-and-place (eyeglasses), and lighter flipping achieved >90% success within 50 training episodes. Generated assets displayed stable, physically plausible behavior in contact-rich regimes, without need for additional tuning.
• Dataset bridging vision, language, and physics: By standardizing geometry and physical parameters, PhysX-Mobility supports research at the intersection of 3D vision, articulated object understanding, robotic manipulation, and policy learning, as well as benchmarking generative frameworks such as PhysX-Anything.

A plausible implication is that PhysX-Mobility can serve as a foundational resource for research in sim-to-real transfer, model-based reinforcement learning with articulated objects, and scalable generative modeling from text or RGB input.

7. Distinction from Related Datasets and Limitations

PhysX-Mobility extends PartNet-Mobility by approximately doubling the category scope and providing richer, simulator-verified physical parameters for all objects (Cao et al., 17 Nov 2025). Unlike datasets focused on visual perception (e.g., ShapeNet), PhysX-Mobility provides kinematic annotations, joint limits, and complete physics blocks in all assets. The use of run-length serialized voxels enables large-scale text-conditioned training regimes via VLMs.

A limitation is that not all parameters (e.g., fine-scale dynamic friction or elastic deformation) are measured directly from physical objects, but rather sampled or derived from canonical databases and analytic calculations. Quality control is partly manual, with >10% of assets undergoing spot-check by annotators for plausible behavior—a common practice in large-scale dataset construction. The physical and kinematic annotations are sufficient for most rigid-body simulation tasks but may require adaptation for domain-specific extensions (e.g., soft body physics or granular manipulation).

PhysX-Mobility fills a key gap in providing richly annotated, articulated, and physical grounding for a diverse range of real-world objects, enabling research at the intersection of vision, language, robotics, and physical simulation (Cao et al., 17 Nov 2025).