RFW-A

RFW-A: Agentic Procedural Generation of 400K Articulated Assets over 500 Categories

Indoor · Articulated assets generated with RFW-AExplore this scene ↓
Read the abstract ＋

IMAGE → GENERATOR → ARTICULATED ASSETS

Explore RFW-A-400K

One category generator. Many possibilities. Explore variations in structure, materials, and motion.

402,410articulated assets
531object categories
2.68Mparsed links
84.65%assets with multiple movable joints

Find a category

Loading image-grounded generators…

Explore scenes

Select an object to explore its motion, mass, materials, and joints.

Indoor scene with selectable objects
Indoor · Select a highlighted object

Select a numbered object in the scene or use the buttons above.

01 / 04Indoor

Method

Build a category generator, validate it, then sample articulated assets with grounded materials and physics.

Generation cost

Adjust the sample count to see the API cost per asset.

11,000 ASSETS
RFW-A$0.053per asset
Per-asset agentic synthesis$1.237per asset

Estimated API cost only; procedural compute is excluded. Mean generator authoring: $10.62. Based on 10 categories and 200 successful baseline synthesis records.

Directed editing

Pick an object and change exactly one thing — a component, a count, a material, a dimension. Every edited model is a real run of the same generator with one field changed; the reference configuration stays fixed for comparison.

1 Pick an object2 Choose what to change3 Compare side by side

BEFORE
drag to orbit
AFTER
drag to orbit
Change
To
CONFIGURATION EDIT

Every output is a real run of the same generator with exactly one field changed — not a re-sampled lookalike. Models are baked in a fixed pose; joint interaction lives in the asset explorer. Cameras share a fixed frame per edit group, so size edits stay true to scale.

Materials & physics

A shared library links PBR appearance to six physical properties. Seeded material selection binds appearance and physical annotations together.

18families
124material classes
8,295PBR materials

A consistent material vocabulary connects appearance and physical behavior across the dataset.

ONE MATERIAL ENTRY → SIX PHYSICAL PROPERTIES
ρDensity kg/m³EYoung’s modulus PaνPoisson’s ratio —μₛStatic friction —μₖDynamic friction —eRestitution —

Physical validation

Kinematic validity, simulator import, and stable rollouts.

KINEMATIC VALIDITY100.0%

Full-release assets passing the kinematic validity check.

SIMULATOR IMPORT100.0%

Core-200 assets imported across the evaluated simulators.

STABLE ROLLOUT99.5%

One-second stable rollouts on Core-200.

Results

Dataset scale, asset quality, and articulation estimation. Select a column to sort.

Qualitative results ＋Qualitative comparison of PM-only, PM plus Articraft, PM plus RFW-A-400K and ground truth

Robotic manipulation

Cabinet

Open the cabinet door with the left arm, then close it with the right arm.

Loading task videos…

Select a video to enlarge.

Real-world success

62%

Retrieval

72%

Retrieval + 4 Edit

80%

Retrieval + 9 Edit

500 simulation demonstrations per task, with the same policy architecture and training budget. Mean success over five tasks, with 10 real-world trials per task.

Results by task

Citation

@misc{rfwa2026,
  title = {RFW-A: Agentic Procedural Generation of 400K Articulated Assets over 500 Categories},
  author = {Lyu, Yuanbo and Pan, Changjian and Zhang, Saining and Yang, Yi and others},
  year = {2026},
  note = {Public identifier to be added}
}

Task video