DeformSmith: Physics Harness-Guided Hierarchical Generation of Deformable Assets for Robot Manipulation
DeformSmith generates deformable robot-manipulation assets by testing and refining their physics, not just their shape.
The paper describes a framework that builds assets from text or a single image, covering geometry, appearance, physical models, and material behavior. A shared physics harness checks how the object deforms and responds to robot interaction, then feeds that evidence back into refinement. The authors report better visual quality and physical plausibility than PhysGen3D, PhysGM, and PhysX-Omni. HF Daily Papers' note
The paper describes a framework that builds assets from text or a single image, covering geometry, appearance, physical models, and material behavior. A shared physics harness checks how the object deforms and responds to robot interaction, then feeds that evidence back into refinement. The authors report better visual quality and physical plausibility than PhysGen3D, PhysGM, and PhysX-Omni. HF Daily Papers' note
score 4