Geometry
Reconstruct and align meshes, particles, and Gaussian appearance.
FROM VISUAL IDEAS TO PHYSICAL INTERACTION
1 Nankai University 2 Rightly Robotics, A4x

MANIPULATION IN THE LOOP
More text-to-manipulation demos coming soon.
Scene and close-up views show the robot handling a generated deformable object.
Qualitative interaction replay. Task acceptance and material consistency are evaluated separately in the paper.

OVERVIEW
Text and images tell us what an object looks like. Interaction reveals how it behaves.
DeformSmith constructs deformable assets from text or a single image. A hierarchy of agents builds geometry, establishes physical models, configures material behavior, and plans robot interaction. A shared physics-grounded harness uses simulation probes and manipulation feedback to guide construction and refinement.
The outputs bring together geometry, appearance, physical parameters, validation records, and replayable interaction data in a common asset representation.
THE METHOD
Progressively construct each part of an asset, then examine its behavior through simulation.
Reconstruct and align meshes, particles, and Gaussian appearance.
Establish mass, volume, contact conditions, and the initial state.
Evaluate material candidates with deformation and recovery probes.
Use pick-and-place feedback to revise actions and material choices.

QUALITATIVE COMPARISON
More demos coming soon.
Explore the same object across four generation methods during a five-second gravity drop.
Loading comparison videos…
REFERENCE
@misc{li2026deformsmith,
title = {{DeformSmith}: Physics Harness-Guided Hierarchical Generation of Deformable Assets for Robot Manipulation},
author = {Li, Can and Gu, Jie and Deng, Zishun and Chen, Jingmin and Sun, Lei},
year = {2026},
eprint = {2609.18620},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.18620}
}