This supplement brings together the saved visual outputs for the method comparison and the language-model ablation.
Two comparisons
Method comparison: PhysX-Anything, URDF-Anything+, ArtLLM, Instruct-Particulate, Procedura, Articraft (Kimi-K3), and MARC (Kimi-K3), across 20 reference objects.
API ablation: Kimi-K3 and GPT-6, each with and without MARC. Kimi-K3 has 20 objects; each GPT-6 set has 19 and does not include the planetary gearset.
Reading the motion
Choose a recorded control from the menu on each card. Play all cards together, drag the shared slider, or scrub one card independently. Static views remain still. Click a render to enlarge it.
- Initial / actuated: two recorded states from the selected comparison input; no intermediate motion is synthesized.
- Native control / physics replay: recorded mechanism-control responses or native simulator body transforms.
- Joint samples / kinematic preview: poses obtained from declared joints or authored animation. A joint-range sweep alone does not establish passive coupling or collision-free motion.
- Static / partial: retained geometry is shown even when articulated export or program execution failed.
Native physics previews preserve their original representation, including MuJoCo XML for PhysX-Anything and USD constraints for Procedura. Collision settings are stated on the corresponding tracks. No new coupling is added for display.
Display conventions
Playback synchronizes normalized sequence progress; it does not equate different methods’ input magnitudes. Individual tracks retain fixed views, but camera framing and materials can differ between archived recordings. The selected comparison renders use aligned views. Rendered output geometry and motion have not been edited with image generation.
Watermarks on reference images were removed for presentation using GPT Image. These display copies do not replace the original inputs used to generate or evaluate the models.
The gallery is a visual supplement, not a replacement for the quantitative evaluation protocol in the paper.