DeformSmith builds deformable objects for robot hands from one image

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to act on yet. The input is an abstract and a project page: no scores, no evaluation date, no released code. If you generate soft-body assets for manipulation training, the three baselines now to compare against are PhysGen3D, PhysGM and PhysX-Omni — wait for numbers before switching.

Posted to arXiv on 16 September 2026, DeformSmith is a system that builds simulated deformable objects — things that bend, squash or stretch when touched — for robot manipulation, starting from a text description or a single image. It generates geometry, material behaviour and robot interaction in stages, checking and refining each stage against a shared physics-based test loop. The authors report it beats PhysGen3D, PhysGM and PhysX-Omni on visual quality and physical plausibility, but the paper gives no scores, no evaluation dates and no numbers.

Key facts

  • ·The paper is arXiv 2609.18620, posted 16 September 2026. source
  • ·DeformSmith generates deformable objects for robot manipulation from text or a single image. source
  • ·It builds geometry, physical models, material behaviour and robot interaction in stages, using a shared physics-grounded harness. source
  • ·The authors report better visual quality and physical plausibility than the baselines PhysGen3D, PhysGM and PhysX-Omni. source
  • ·The abstract gives no scores, no evaluation dates and no quantitative comparison against those baselines. source
  • ·A project page is listed at https://can-lee.github.io/deformsmith-web/. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire