--- pretty_name: UniPhys-Bench license: cc-by-nc-4.0 language: - en tags: - 3d - physical-grounding - articulation - robotics - simulation - benchmark size_categories: - 1K
**Redder = more interactive** | All articulated parts in motion | | Example 1 original mesh | Example 1 part decomposition | Example 1 affordance visualization | Example 1 articulated motion | | Example 2 original mesh | Example 2 part decomposition | Example 2 affordance visualization | Example 2 articulated motion | | Example 3 original mesh | Example 3 part decomposition | Example 3 affordance visualization | Example 3 articulated motion | ## Simulation-based embodied interaction The following benchmark-level demonstrations illustrate downstream embodied interaction with assets from the complete UniPhys-Bench. **1. Drawer Pulling** | Initial State → | Physical Contact → | Actuated State | | :---: | :---: | :---: | | Drawer pulling initial state | Drawer pulling physical contact | Drawer pulling actuated state | **2. Faucet Turning** | Initial State → | Physical Contact → | Actuated State | | :---: | :---: | :---: | | Faucet turning initial state | Faucet turning physical contact | Faucet turning actuated state | **3. Laptop Closing** | Initial State → | Physical Contact → | Actuated State | | :---: | :---: | :---: | | Laptop closing initial state | Laptop closing physical contact | Laptop closing actuated state | ## Benchmark scope and construction | Statistic | Value | | --- | ---: | | Articulated objects | 1,927 | | Parts | ~16K | | Motion-relevant components | ~5.5K | | Entity ID range | `UPB_00000000`–`UPB_00001926` | | Role | Evaluation-only benchmark | The benchmark combines 1,473 articulated 3D objects from a commercial asset provider, whose identity is omitted for review and whose part decompositions were created by professional designers, with 454 evaluation-only objects held out from the source distribution used to construct UniPhys-40K. All objects use the same annotation schema and evaluation protocol. Entity IDs are globally unique and must not be renumbered. The public source collection draws on [Objaverse-Sketchfab](https://objaverse.allenai.org/), [HSSD](https://3dlg-hcvc.github.io/hssd/), [3D-FUTURE](https://github.com/3D-FRONT-FUTURE/3D-FUTURE-ToolBox), [ABO](https://amazon-berkeley-objects.s3.amazonaws.com/index.html), and [PartNet](https://partnet.cs.stanford.edu/). Its curation uses [TRELLIS-500K](https://github.com/microsoft/TRELLIS/blob/main/DATASET.md) quality-filtered asset lists for the first four sources and aligns PartNet components with [ShapeNet](https://shapenet.org/) meshes to recover appearance. These third-party references describe upstream data sources. The UniPhys pipeline first produces articulation and physical-property annotations. Human annotators then inspect and correct joint type, axis, pivot, motion range, motion-dependent part groupings, and intrinsic physical-property plausibility for motion-relevant components. The benchmark covers diverse object categories, structural complexities, part granularities, modeling styles, scales, masses, materials, and articulation patterns. ## Annotation scope | Task | Inputs | Ground truth | | --- | --- | --- | | Part-level intrinsic physical grounding | object and target-part geometry | part identity, semantic descriptions, material, density, Young's modulus, hardness, Poisson's ratio, friction, graspability, and affordance | | Kinematic parameter grounding | object and target-part geometry | prismatic/revolute joint type, axis, pivot, and motion range | | Articulation structure grounding | object and target-part geometry plus candidate-part metadata | IDs of parts that move together with the target part | | Object-level physical grounding | complete-object geometry | object identity, category, dimensions, and mass | Part-property units follow the paper: density is in `g/cm^3`, Young's modulus in `GPa`, hardness in `HV`, and Poisson's ratio and friction are unitless. Object dimensions are `[L, W, H]` in centimeters and mass is in kilograms. Affordance is scored from 1 to 10, with smaller values indicating higher affordance. Motion labels use `B` for prismatic translation and `C` for revolute rotation. The broader part-level annotations additionally use `A` for contact-only and `D` for rigid or fixed parts. ## Dataset structure Each benchmark object is stored in one `UPB_` directory. For example: ```text UPB_00000000/ ├── annotations/ │ ├── object.json │ ├── part_2.json │ ├── part_3.json │ └── ... ├── full_model/ │ ├── model.obj │ ├── material.mtl │ └── texture files ├── meta_data.json ├── model.urdf ├── parts/ │ ├── part_2.obj │ ├── part_3.obj │ ├── material and texture files │ └── ... └── plys/ ├── model.ply ├── 2.ply ├── 3.ply └── ... ``` | Path | Description | | --- | --- | | `parts/part_.obj` | Decomposed part mesh, with its available MTL and texture assets. | | `full_model/model.obj` | Complete object OBJ produced by concatenating all released part meshes. | | `plys/model.ply` | Point cloud for the complete object. | | `plys/.ply` | Part point cloud aligned with `parts/part_.obj`. | | `annotations/object.json` | Object identity, category, real-world dimensions, and mass. | | `annotations/part_.json` | Part semantics, intrinsic physical properties, articulation parameters, and dependency group. | | `meta_data.json` | Entity provenance, object summary, geometry metadata, and annotation version. | | `model.urdf` | Articulated assembly of all released parts using the annotated joints and motion parameters. | The same numeric part ID is used by `part_.obj`, `.ply`, and `part_.json`, so geometry and annotations can be joined without an additional mapping file. Part IDs need not start at 1 or be contiguous. ### Simulation-ready URDF `model.urdf` assembles all parts and encodes the annotated joint types and motion parameters. During export, the mesh is rescaled using the annotated object dimensions so that the assembled asset has a real physical size. URDF geometry uses **meters**, while `annotations/object.json` stores dimensions in centimeters. > [!IMPORTANT] > The released URDF provides the articulated geometry and kinematic assembly, > but intrinsic physical properties are not written into the URDF. Density, > friction, mass, and other physical values should be read from the JSON > annotations and assigned as needed for the target simulator and experiment. ### Metadata `meta_data.json` stores the entity's provenance, object summary, geometry metadata, and annotation version. The `source` field may be a list of provenance records or a single record; readers should handle both forms. The following schema example uses a list. The provider name is anonymized for review: ```json { "id": "UPB_00000454", "source": [ { "dataset": "UniPhys-Bench", "organization": "Anonymous asset provider", "license_ref": "UniPhys-Bench" } ], "object": { "category": "Furniture/StorageFurniture", "object_name": "Corner Display Cabinet with Drawer" }, "geometry": { "asset_type": "decomposed_parts", "num_parts": 6, "format": "obj" }, "annotation": { "version": "v1.0" } } ``` A public-source entity may instead store a single provenance record, including its upstream asset identifier: ```json { "id": "UPB_00000000", "source": { "dataset": "ABO", "original_id": "B00XBC3BF0", "license_ref": "ABO" }, "object": { "category": "Furniture/SeatingFurniture", "object_name": "Executive Office Chair" }, "geometry": { "asset_type": "decomposed_parts", "num_parts": 18, "format": "obj" }, "annotation": { "version": "v1.0" } } ``` The public-source dataset name and original asset ID describe upstream provenance, rather than the benchmark authors. Provider-identifying fields in review copies should use anonymous labels, as illustrated above. ### Part-level annotations Each `annotations/part_.json` contains the part description, intrinsic physical properties, motion type, joint parameters, and motion dependency: ```json { "index": { "type_name": "default", "entity": "UPB_00000000", "label": "3" }, "part_level": { "part_name": "Front Seat Trim Bar", "affordance": 2, "graspable": false, "basic_description": "Slim ABS trim piece at the front edge of the seat.", "functional_description": "Covers a seam and provides a finished edge.", "movement_description": "Revolute; attached to the seat base.", "grasp_description": "Grasp the handle of the bar." }, "basic_info": { "material": "metal/Steel", "density": 7.85, "young": 200.0, "hardness": 180.0, "poisson": 0.3, "friction": 0.45 }, "kinematic_info": { "motion_types": ["C"], "motion_info": { "dependency": [3], "C": { "axis": [-1.0, 0.0, 0.0], "pos": [-0.02277967, 0.38144422, -0.02326505], "range": [0.0, 0.785], "damping": 0.03 } } } } ``` `kinematic_info.motion_info.dependency` lists the part IDs that move together. For movable parts, `B` contains prismatic parameters and `C` contains revolute parameters; `axis`, `pos`, `range`, and `damping` describe the corresponding joint. `range` is the `[lower, upper]` motion interval; for a revolute (`C`) joint, both limits are rotation angles expressed in radians. ### Object-level annotations `annotations/object.json` contains the object identity and global physical properties: ```json { "object_name": "Example Object", "category": "Furniture/StorageFurniture", "volume": [70.0, 68.0, 110.0], "mass": 15.5 } ``` `volume` is `[length, width, height]` in centimeters and `mass` is in kilograms. The values above illustrate the schema; released files contain the verified values for each entity. ## Access and preparation Use the dataset and code supplied through the anonymous submission materials. Public author and organization accounts are omitted from this review README. Extract all 1,927 object directories into one data root, preserving the original entity IDs: ```text data/UniPhys-Bench/ ├── UPB_00000000/ ├── UPB_00000001/ ├── ... └── UPB_00001926/ ``` Run the following commands from the root of the accompanying code package. Generate model-ready point clouds: ```bash python pre_process/generate_npzs.py \ --data_root data/UniPhys-Bench \ --output_dir data/UniPhys-Bench-processed/npzs ``` Generate one inference manifest for each evaluation task: ```bash python pre_process/generate_jsons_for_inference.py \ --data_root data/UniPhys-Bench \ --npz_dir data/UniPhys-Bench-processed/npzs \ --output_dir data/UniPhys-Bench-processed/manifests ``` The command writes: ```text manifests/ ├── intrinsic_physics_part.json ├── intrinsic_physics_object.json ├── kinematic_parameters.json └── articulation_structure.json ``` Each sample keeps its complete benchmark annotation. UniPhysGen inference embeds that record as `source_sample`, which the evaluation package reads as ground truth. Prepare and evaluate all 1,927 objects together using this single data root and the four generated manifests. The complete benchmark uses one evaluation protocol; no additional dataset-specific preparation or separate evaluation is required. ## Evaluation protocol The paper reports the following metrics: | Category | Metrics | | --- | --- | | Kinematic parameters | joint-type accuracy, axis angular error in degrees, pivot-to-ground-truth-axis distance, and motion-range mIoU | | Articulation structure | set mIoU and micro-F1 over motion-coupled part IDs | | Material properties | material-category accuracy, density ALDE, and friction MAE | | Object scale and mass | ALDE and MnRE for dimensions and mass | | Affordance | MAE over the 1-10 affordance score | Kinematic parameters are evaluated on ground-truth movable parts, separating parameter estimation from movable-part identification. Axis directions `a` and `-a` are treated as the same articulation axis. Pivot error is measured in a shared AABB-normalized object frame as point-to-ground-truth-axis distance. Motion ranges are canonicalized to unsigned intervals before computing IoU. Run the four evaluators on prediction JSON files or directories of per-sample records: ```bash python -m eval intrinsic_physics_part PREDICTIONS \ --output intrinsic_physics_part_metrics.json python -m eval intrinsic_physics_object PREDICTIONS \ --output intrinsic_physics_object_metrics.json python -m eval kinematic_parameters PREDICTIONS \ --output kinematic_parameters_metrics.json python -m eval articulation_structure PREDICTIONS \ --output articulation_structure_metrics.json ``` Refer to the README in the accompanying anonymous code package for the complete inference workflow and checkpoint commands. ## Intended use UniPhys-Bench is intended for research evaluation of unified physical grounding, including articulation reasoning, physical-property estimation, simulation-ready asset construction, embodied AI, and robotics simulation. It is an evaluation benchmark and should not be mixed into UniPhysGen training or model-selection data. ## Scope and usage considerations - Human annotators verify and correct articulation annotations and assess physical-property plausibility to support consistent research evaluation. Physical values are reference estimates for the depicted objects. - The benchmark cannot cover every object category, material, mechanism, part granularity, or mesh failure mode. - Results can depend on geometric completeness, scale correctness, texture quality, and candidate part decomposition. ## Licensing and provenance The complete contents of the benchmark—including assets, annotations, metadata, derived representations, and dataset organization—are licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Commercial use is not permitted. The full license terms and any applicable upstream notices remain in effect. Use the `source` field in each entity's `meta_data.json` to retain provenance. For a list of records, inspect each entry; for a single record, inspect that object directly. Applicable fields include `dataset`, `original_id`, `organization`, and `license_ref`. Provider names are anonymized in this README's review examples. This does not change license terms or attribution obligations. Retain required third-party creator credits, license notices, and modification notices when sharing data. UniPhysGen source code and model weights are licensed separately; this dataset license does not replace their licenses.