A vision-language agent proposes semantic parts and physical priors while a pretrained 3D generator creates a textured object. The workflow constructs bindings between geometry and mechanisms, observes rendered or simulated behavior, and makes targeted refinements using geometric tools and execution measurements.
PhysAg3D is useful for research into interactive asset creation and agent-assisted 3D authoring. The released controller, preparation tools, worker protocol, and validation references document how to assemble a configured workflow and retain evidence of what an exported asset can actually do.
