PhysX-Mobility: Simulation-Ready 3D Assets

• PhysX-Mobility is a comprehensive dataset of simulation-ready 3D assets with detailed geometry, physical properties, and articulated mechanisms.
• The dataset uses rigorous cleaning, automated manifold checks, voxel-based segmentation, and synthetic augmentation to ensure high annotation quality.
• Optimized for simulators like MuJoCo, it supports advanced physics tasks, embodied AI, and robotic policy learning with impressive performance metrics.

PhysX-Mobility defines a large-scale dataset of richly annotated, simulation-ready 3D assets aimed at advancing physics-based simulation, embodied AI, and robotics research. The dataset substantially expands the semantic diversity and annotation richness beyond previous physical 3D benchmarks, offering over two thousand high-fidelity models with explicit geometry, physical properties, and articulated mechanisms, all optimized for direct integration with popular simulators such as MuJoCo. PhysX-Mobility is a cornerstone resource within the "PhysX-Anything" framework, facilitating simulation-ready asset generation and rigorous evaluation on contact-rich tasks (Cao et al., 17 Nov 2025).

1. Dataset Construction and Scale

PhysX-Mobility aggregates and extends previous repositories by applying rigorous cleaning, annotation, and synthetic augmentation:

• Source models include approximately 1,200 PartNet-Mobility scans and meshes, supplemented by 1,000 additional consumer-grade RGB-D scans. Each scan is paired with multiple real-world photographs for visual validation.
• Annotation pipeline: Automated manifold checks and manual mesh repair are performed. Rigid part segmentation is realized via voxel-based tagging. Material classes are assigned by database lookup; physical attributes (density, mass, inertia) are calculated analytically.
• Articulation extraction: Kinematic graphs are constructed from PartNet-Mobility, with manual axis and limit verification.
• Synthetic augmentation: Achieved via uniform random scaling (in [0.9, 1.1] for 20% of instances), friction coefficient jittering (±5%), and generation of 100 "hybrid" assemblies by swapping parts.
• Scale and coverage: The dataset comprises 2,184 unique 3D assets, expanding category coverage by over 2× relative to prior datasets. There are 47 semantic categories, including furniture (800 instances), office equipment (210), kitchen appliances (230), electronics (310), tools (180), household fixtures (240), and toys/miscellaneous items (214). Category sizes are maintained within ±20% variability.

2. Physical and Articulation Annotations

Each object decomposition features meticulous labeling of physical and mechanical properties:

• Physical properties per rigid part:
  • Mass $m = \rho V$, with $\rho$ the database-assigned density and $V$ part volume (from mesh).
  • Center of mass $\mathbf{c} \in \mathbb{R}^3$.
  • Inertia tensor $\mathbf{I} \in \mathbb{R}^{3 \times 3}$, diagonalized in the local frame.
  • Friction (static $\mu_s$, dynamic $\mu_d$) and restitution coefficient $e$.
• Articulation semantics:
  • Joint type $$J \in \{\text{revolute}, \text{prismatic}, \text{ball}\}$$.
  • Axis direction $\hat{\mathbf{u}} \in \mathbb{R}^3$, anchor $\mathbf{p} \in \mathbb{R}^3$.
  • Joint limits $[\theta_{\min}, \theta_{\max}]$ (revolute) or $[d_{\min}, d_{\max}]$ (prismatic).
  • Degrees of freedom: 1 for revolute/prismatic, 3 for ball joints.
  • For 15% of instances: Optional linear spring–damper with stiffness $k$ and damping $b$, and actuation torque limit $|\tau| \leq \tau_{\max}$.

3. 3D Representation, Tokenization, and Compression

PhysX-Mobility pioneers a compact, simulation-efficient 3D data structure:

• Coarse geometry: Occupancy grids on a $32^3$ voxel lattice ($\mathbf{V}^{\rm low} \in \{0,1\}^{32^3}$).
• Sparse serialization: Occupied voxel indices are linearized and merged into ranges, then serialized as a comma-separated string.
• Compression: This method achieves a $\sim$193× token count reduction relative to naïve mesh serialization (mean $\sim$26 tokens/asset), optimizing compatibility with vision-LLM (VLM) context windows and efficient downstream processing.
• Fine-grained decoding: A flow-based transformer refines low-res voxels to $128^3$, supporting high-fidelity meshing for simulation use.

4. Data Format and Simulator Integration

Data is distributed for ready loading into major physics engines without manual post-processing:

• Organization and schemas:
  • model.urdf for kinematic and physical specification.
  • model.sdf for MuJoCo-style environments.
  • Per-part mesh (PLY/OBJ), texture (PNG/JPEG), and a metadata.json summary.
• Simulator integration: All physical and articulation properties, including joint limits, damping, friction, and inertia, are encoded natively in URDF/SDF and accessible via standard APIs (e.g., mujoco_py).
• Example pipeline: A MuJoCo-compatible object can be loaded and rendered interactively with a few lines of Python, with all physical parameters faithfully preserved from annotation to runtime.

5. Dataset Statistics and Category Distribution

Summary measures of PhysX-Mobility’s scale and diversity include:

Statistic	Value/Range
Object count	2,184
Categories	47
Longest dim (mean, $\sigma$)	0.45 m, 0.32 m
Parts/object (mean, $\sigma$)	3.1, 1.4
Mass (mean, range)	3.2 kg (0.05–45 kg), $\sigma$ = 5.7 kg
Static friction $\mu_s$ (mean, $\sigma$)	0.42, 0.15
Dynamic friction $\mu_d$ (mean, $\sigma$)	0.35, 0.12
Articulation complexity	1,428 fixed, 512 1-DOF, 244 multi-DOF objects

Category distribution examples:

Group	Example Objects	Count
Furniture	Chair, table, sofa, bed	800
Office Equipment	Stapler, monitor	210
Kitchen	Coffee machine, toaster	230
Electronics	Camera, fan	310
Tools	Screwdriver, drill	180
Fixtures	Faucet, cabinet door	240
Toys/Misc	Toy car, puzzle box	214

6. Generative Benchmarks and Downstream Performance

Performance on generative and robotic manipulation tasks is established by direct comparison to state-of-the-art baselines:

• Generative 3D asset benchmarks (higher better for PSNR/F-Score/material/affordance/kinematic/description; lower better for Chamfer, scale error):

Metric	PhysX-Anything	PhysXGen	Articulate-Anything	URDFormer
PSNR (↑)	20.35	20.33	16.90	7.97
Chamfer CD (↓)	14.43	14.55	17.01	48.44
F-Score (↑)	77.50	76.30	67.35	43.81
Abs. scale error (↓)	0.30	43.44	—	—
Material acc. (↑)	17.52	6.29	—	—
Affordance acc. (↑)	14.28	9.75	—	—
Kinematic param (↑)	0.94	0.71	0.65	0.31
Description coher. (↑)	19.36	12.89	—	—

• Robotic policy learning: In a MuJoCo-style simulation over 120 manipulation tasks (e.g., turning faucets, opening cabinets), policies trained on PhysX-Mobility assets achieved a 92% success rate and normalized reward of 0.78, compared to 61% and 0.45 for the PhysXGen baseline. This demonstrates not only generative fidelity but immediate usability for contact-rich, articulated interactions.

7. Context, Applications, and Impact

PhysX-Mobility enables simulation-ready, richly annotated physical assets for:

• Embodied AI: Facilitating generalization and policy learning in robotic manipulation, tool use, and interactive perception.
• Computer vision and 3D generation: Serving as a primary resource for VLM-based geometry-aware modeling, supporting advanced tokenization and context-efficient training protocols.
• Physics-based simulation: Providing a comprehensive testbed for benchmarking sim-to-real transfer, active safety evaluation, and articulated object control in environments including MuJoCo and other URDF/SDF-compatible engines.
• Dataset design: Demonstrating annotation, compression, and integration best practices, including explicit calculation of mass, inertia, and kinematic limits directly from mesh and physical database lookups.

A plausible implication is that PhysX-Mobility’s design—specifically its 193× geometry token compression, granularity of ground-truth physics metadata, and extensive semantic coverage—will serve as a reference standard for simulation-ready physical datasets in robotics and AI for the foreseeable future (Cao et al., 17 Nov 2025).