MPArt-90: Articulated Object Benchmark

• MPArt-90 is a large-scale, multi-part articulated object benchmark featuring multi-view RGB-D data and precise ground-truth articulation parameters.
• It overcomes prior limitations by integrating diverse objects, detailed part segmentation, and dual articulation states for robust evaluation.
• The benchmark employs rigorous metrics like Axis Angle Error and Chamfer Distance to advance practical applications in simulation and robotic interaction.

MPArt-90 is a large-scale, multi-category, multi-part articulated object benchmark devised to rigorously evaluate methods for joint geometry-motion reconstruction of articulated objects. Designed as the central benchmark in "GaussianArt: Unified Modeling of Geometry and Motion for Articulated Objects," MPArt-90 addresses critical scalability and generalization gaps in prior benchmarks by introducing multi-view, multi-state RGB-D data and ground truth articulation for a diverse set of objects spanning a wide range of complexity (Shen et al., 20 Aug 2025).

1. Motivation and Limitations of Prior Articulated Benchmarks

Prior articulated object reconstruction datasets exhibit several core deficiencies: limited object count (often fewer than 20), narrow articulation diversity, and a predominance of simple two- or three-part objects. These constraints prevent failure cases in scalability, robustness, and generalization from becoming apparent. Furthermore, many earlier datasets lacked full multi-view, multi-state ground truth, hampering physically grounded evaluations. MPArt-90 was specifically created to overcome these barriers and systematically expose weaknesses in methods reliant on decoupled geometry and motion pipelines, fragile clustering, or brittle initialization routines. The benchmark's stated aim is to enable quantitative and qualitative assessment of articulated object reconstruction methods under realistic, challenging articulation complexity.

2. Benchmark Construction and Composition

MPArt-90 comprises 90 articulated objects drawn from 20 object categories. The benchmark predominantly consists of synthetic objects (87 from PartNet-Mobility) and includes 3 real-world objects sourced from MultiScan, prioritized for available high-quality internal structure annotation. Additionally, 12 objects from the PARIS Two-Part dataset and 2 from DigitalTwinArt-PM are included within the synthetic set, thus ensuring both diversity and some continuity with past benchmarks. Each object features high-fidelity, ground-truth annotation of articulation parameters and is rendered in two distinct articulation states with dense multi-view RGB-D coverage (100 rendered training views, 20 test views per state, images at 800×800 resolution).

The benchmark includes objects spanning from 2 to 20 parts, with explicit stratification for performance analysis:

Part Count	Number of Objects
2	34
3	21
4–5	24
6–20	11

Types of articulations include both revolute and prismatic joints, with each object's articulation configuration sampled to ensure diverse, nontrivial combinations.

3. Dataset Acquisition, Annotation, and Articulation Representation

Synthetic objects are rendered using BlenderProc2 from PartNet-Mobility asset geometry, with two states chosen by random sampling in specified canonical and extreme ranges ([0.65, 0.75] and [0.35, 0.45], respectively) to maximize inter-state visibility and articulation diversity. Real objects are selected from MultiScan for annotation reliability, as annotating hidden articulation parameters remains infeasible in most real-world scanned datasets.

Each object instance provides:

• Full part segmentation masks for all articulated components.
• Canonical-to-current-state ground-truth rigid transformations and joint parameters.
• Articulation is parameterized as per-part rigid transformations, with ground-truth evaluation measuring correspondence against these canonical motions.

All benchmarks use a 3D Gaussian-based representation to assign geometry and part identification; each Gaussian encodes geometry and soft/hard affiliation to one of $N$ parts. Articulated motion is represented via per-part rigid transforms acting on the constituent Gaussians.

4. Evaluation Protocol and Metrics

For each object and state, a multi-view split is adopted: 100 views for training, 20 for testing. All results are reported in the high-visibility state and averaged across three random seeds.

Evaluation metrics include:

• Axis Angle Error: Angular difference between estimated and true joint axis direction.
• Axis Pos Error: Euclidean distance between estimated and ground-truth joint axis origins.
• Part Motion Error: Difference in joint motion parameter (degrees for revolute, meters for prismatic).
• CD-s (Chamfer Distance – static parts): Chamfer distance between recovered and ground-truth meshes for static parts, using 10,000-point uniform samplings.
• CD-m (Chamfer Distance – moving parts): Same as CD-s, but restricted to dynamic parts.

Benchmark results are stratified by part count, revealing algorithmic robustness or degradation with increasing articulation complexity.

5. Relationship to GaussianArt and Comparative Methodology

MPArt-90 is directly coupled with the GaussianArt reconstruction framework, which embodies a unified, articulated-3D-Gaussian representation for joint geometry and motion estimation. The benchmark serves to expose the failure modes of prior methods—particularly cluster-based and decoupled geometry-motion pipelines—when confronted with increased part count and intricate articulation. GaussianArt leverages robust initialization (with fine-tuned part mask segmentation, multi-view matching, and reprojection), soft-to-hard assignment optimization, and joint geometry-motion learning. This allows MPArt-90 to be used systematically for ablation studies, demonstrating that performance collapses without robust segmentation and initialization strategies as part count grows.

Notably, quantitative comparisons show that while ArtGS performs comparably on 2-part objects, its errors (especially CD-m) escalate rapidly with higher part counts, while GaussianArt maintains consistent, moderate error across the stratified spectrum.

6. Practical Implications and Downstream Applications

MPArt-90 has several key practical implications:

• By including large, compositional, multi-part objects, it demonstrates the necessity of unified geometry-motion modeling for articulated reconstruction, moving beyond the simple two-component targets ubiquitous in earlier work.
• Outputs generated from methods benchmarked on MPArt-90 are directly convertible to physics-based simulation environments (e.g., NVIDIA Omniverse IsaacSim) for robotic interaction with digital twins, enabling studies of manipulation, dynamic planning, and embodied AI.
• The canonical articulated representations are suitable for integration with 4D human-scene modeling, scene synthesis, and animation pipelines, as the provided ground-truth supports dynamic scene creation with complex object interactions.

7. Significance for Scalability and Generalization in Articulated Object Research

MPArt-90 exposes critical limitations of previous approaches by including challenging object instances with up to 20 articulated parts, ensuring that only methods designed for true scalability and generalization remain robust. Its design paradigm—density, diversity, and multi-state multi-view data with explicit ground-truth—motivates research not just in object-centric reconstruction, but also in the automated construction of digital twins, simulation-based policy learning, and adaptive physical modeling pipelines. Empirical analyses and ablations underscore that scalable multi-part articulated reconstruction is only possible with robust, joint representations and careful part initialization; decoupled or clustering-based approaches rapidly degrade beyond 2–3 parts (Shen et al., 20 Aug 2025).

In summary, MPArt-90 constitutes the largest and most challenging articulated-object reconstruction benchmark to date, enabling systematic, scalable, and physically meaningful evaluation of both existing and emerging methods for joint geometry and motion learning.