DeformSmith generates physics-grounded deformable assets for robot manipulation

Can Li · hf · 2026-09-21

DeformSmith automatically creates interactive, physically plausible deformable assets from text or a single image for robot manipulation—addressing the fact that deformation behavior can't be inferred from text/images alone yet directly affects interaction suitability. It uses hierarchical agentic construction with a shared physics harness to iteratively build, test and refine geometry, physical models, materials and robot interaction, with manipulation feedback closing the loop. It outperforms PhysGen3D, PhysGM and PhysX-Omni on visual quality and physical plausibility while synthesizing manipulation training data.

Original post →

More from Embodied

Embodied channel →