Agents build reusable generators for articulated objects
RFW-A: Agentic Procedural Generation of 400K Articulated Assets over 500 Categories · arXiv:2610.40333
In plain words
The paper introduces a way for software agents to build reusable programs for making movable 3D objects, such as cabinets with doors. Instead of designing each object separately, the program can make and edit many versions by changing its settings.
The released collection has more than 400,000 assets, and every asset passed the study's joint-motion check. In five robot tasks, controllers trained in simulation had higher average success when training used edited versions of the objects. Those results cover the tests the researchers ran, not every object or task.
The research
Can a system make many articulated objects without requiring a person to author a new program for each category or agents to rebuild each asset? Such objects need both part structure and joint motion, and producing those details by hand takes substantial effort. The paper tests whether agents can build reusable generators that scale this work.
RFW-A uses agents to build and validate reusable category-level generators, then runs those programs to create articulated assets without further model calls.
RFW-A first builds category-level programs from reusable components and their assembly and joint rules. Agents test and repair each generator across sampled configurations and motion states before release. The released program then makes assets by execution, without further model calls.
Evaluators found that all 402,410 released assets passed the joint-motion test. In a separate downstream comparison, a model trained with RFW-A assets scored higher on part-overlap measures than one trained with Articraft on both tested splits.
That supports usefulness for the tested estimation task, though one whole-object measure on unseen categories favored Articraft.
These results cover sampled generator configurations, a single articulation-estimation model family, and five real-world manipulation tasks. The tests do not establish accuracy for real-object-motion or transfer to independently sourced objects.
What the evidence shows
- Move from assets to programs
RoboFlywheel-Articulation (RFW-A) is a framework in which software agents build reusable programs for generating articulated 3D objects, such as cabinets with movable doors. The research question is how to scale production when procedural programs require authoring and per-object agentic synthesis repeats reasoning and correction for every asset. RFW-A moves the agent work from individual assets to generators shared by an object category.
- Build and check generators
RFW-A builds a blueprint from reusable parts and rules for how they fit together and move. Agents turn it into executable code, then test and repair it across sampled configurations and motion states.
Once released, the generator can produce additional assets from different random seeds without further model calls. The approach makes reuse possible, while the outputs remain bounded by the components and rules in each blueprint.
- Check assets at two scales
The released RFW-A-400K collection contains 402,410 assets across 531 categories. Evaluators checked every asset for valid joint motion; an asset passed only if all its movable joints passed the 21-position test.
All assets passed. A separate sample of 200 assets was tested in four simulators, and 99.5% passed the one-second fixed-root rollout in all four. These results address different checks and populations.
- Test a downstream model
A downstream test trained the Particulate-B articulation-estimation model on PartNet-Mobility alone or with an additional dataset, then compared predictions on shared evaluation splits. The RFW-A-trained model had stronger part-overlap scores than the Articraft-trained model on both the training-seen split and a split of categories excluded from training.
On the latter split, however, whole-object point-cloud distance was slightly worse, so the gains did not extend to every measure.
- Try edits on real objects
Researchers also trained robot controllers in simulation using retrieved assets alone or with edited variants, then tested them on five real-world tasks. Average success rose from 6.2 to 8.0 trials out of 10 per task when training included nine edited variants, with the double-door task unchanged.
This supports a benefit in the tested tasks, while the small set of tasks and trials limits broader conclusions.
Deep
- Original paper authors
- Yuanbo Lyu, Changjian Pan, Saining Zhang, Yi Yang, Peishuo Li
- Submitted to arXiv
- License
- arXiv non-exclusive