Physical semantics (articulation and intrinsic properties) are critical for realistic robot interaction and embodied AI, yet most 3D asset collections lack unified, verified annotations. This benchmark supplies a focused, human-verified testbed to measure how well models recover kinematic structure, joint parameters, material and density estimates, object scale/mass, and affordance — bridging geometry-only datasets and simulation-ready assets.
What Sets It Apart
- Human-verified motion semantics at benchmark scale: 1,927 articulated objects across two releases (1,473 primary + 454 held-out), ~16K parts and ~5.5K motion-relevant components, with professional designer part decompositions for a large subset. This means evaluation stresses both part granularity and real-world modeling diversity.
- End-to-end simulation readiness: releases include per-entity URDF assemblies (kinematics encoded), per-part and per-object JSON annotations (material, density in g/cm³, Young's modulus in GPa, hardness in HV, Poisson's ratio, friction, affordance scores), and aligned point clouds — enabling direct import into robot simulators after assigning physical parameters from annotations.
- Rigorous, task-specific evaluators: built-in evaluation scripts measure joint-type accuracy, axis angular error, pivot distance, motion-range IoU, articulation-structure mIoU/F1, material-category accuracy, density ALDE, friction MAE, object dimension/mass ALDE and MnRE, and affordance MAE — letting researchers compare models on concrete, simulation-relevant metrics.
Who It's For & Tradeoffs
Great fit if you develop or evaluate models that infer articulation semantics or part-level physical parameters for robotics, embodied AI, or simulator asset pipelines. The dataset is curated for evaluation (not training) and intentionally paired with UniPhys-40K for training experiments. Look elsewhere if you need permissive commercial licensing (this release is CC BY-NC 4.0) or exhaustive coverage of every material/mechanism — the benchmark emphasizes variety and verification but cannot represent every real-world failure mode or every material nuance.
Practical notes
- The two releases share a unified annotation schema and global entity IDs; combine both to reproduce reported paper results. URDF geometry uses meters while object dimension annotations are stored in centimeters. Intrinsic physical values in annotations are reference estimates for simulation assignment and should be validated for high-precision physical experiments.
- Intended uses: evaluation of unified physical grounding (articulation + intrinsic properties), producing simulation-ready assets for downstream robot interaction research, and benchmarking model generalization across heterogeneous part decompositions.
License and provenance: assets and annotations in this release are distributed under CC BY-NC 4.0 (non-commercial); many entities are provided by Manycore Tech with designer-provided decompositions and UniPhys pipeline + human verification applied to annotations.
