BendTwin: Robust Dense-to-Sparse Physical Reconstruction with Bending-Aware Differentiable Spring-Mass Models

NeurIPS 2026 Workshop (PUDM)Oral

Yixiong Jing^1,*, Qi Wang^2,*, Lin Chen³, Junwei Jiang⁴, Guangming Wang^1,†, Haibing Wu¹, Olaf Wysocki¹, Wanli Ma¹, Brian Sheil¹

¹ University of Cambridge ² Institute of Automation, Chinese Academy of Sciences ³ Northwestern Polytechnical University ⁴ The Hong Kong Polytechnic University

* Equal contribution · † Corresponding author

Paper arXiv Code BibTeX

From observation to a physical twin

Reconstruct. Simulate. Predict.

FIG. 01

BendTwin overview: three RGB-D camera views of a sloth toy, its bending-aware spring-mass representation, and simulated deformation results. — Sparse-view RGB-D observations become a deformable digital twin through a bending-aware spring–mass model. View full resolution ↗

19.0%

Lower reconstruction
Chamfer distance

16.8%

Lower future-prediction
Chamfer distance

Graph sampling ratio
tested in sparse ablations

Error reductions compare BendTwin with PhysTwin on the paper’s non-cloth evaluation set (Table 1). Sampling ratio is from Table 2.

01 Construct the twin

The idea

Give sparse models a sense of bending.

Abstract

Reconstructing deformable objects from video requires a model of how they move, not just how they look. Conventional spring–mass systems constrain distances between pairs of nodes, but can lose local shape stability as the physical graph becomes sparse.

BendTwin adds bending stiffness and damping over local surface triplets. By penalizing deviations from rest angles and damping angular motion, it improves reconstruction and future prediction while retaining a differentiable spring–mass formulation. Experiments and ablations evaluate these constraints under graph downsampling and in a surface-only configuration.

BendTwin pipeline: reconstruct surface points from RGB-D observations, add axial springs and bending triplets, optimize geometry and tracking losses, and run differentiable explicit simulation. — From surface reconstruction to mechanical parameter optimization and differentiable simulation. View full resolution ↗

01 / Represent

Local bending triplets

Surface nodes carry axial springs and angular constraints. Bending stiffness resists changes from each triplet’s rest angle.

02 / Optimize

Learn from motion

Geometry and tracking losses fit the mechanical parameters to observed motion. Bending damping dissipates angular oscillation.

03 / Predict

Simulate forward

A differentiable explicit Euler simulator rolls the fitted model forward, from observed reconstruction to unseen future frames.

02 Verify the twin

Experiments & evidence

Better geometry. More faithful motion.

Evaluated on non-cloth deformable objects from the PhysTwin benchmark. The first 70% of each sequence is used for reconstruction and re-simulation; the remaining 30% tests future prediction.

Qualitative reconstruction and future-prediction comparisons between ground truth, PhysTwin, and BendTwin. Red dashed circles highlight differences in local object deformation. — Qualitative comparison of reconstruction, re-simulation, and future prediction. Red dashed circles highlight differences in local object deformation. View full resolution ↗

Quantitative comparison

Paper · Table 1 ↗

Reconstruction & re-simulation
Method	CD ↓	Track ↓	IoU (%) ↑	PSNR ↑	SSIM ↑	LPIPS ↓
PhysTwin	0.0058	0.0091	79.5	28.429	0.963	0.022
BendTwin Ours	0.0047	0.0078	80.6	28.605	0.963	0.021

Future prediction
Method	CD ↓	Track ↓	IoU (%) ↑	PSNR ↑	SSIM ↑	LPIPS ↓
PhysTwin	0.0095	0.0166	68.7	26.412	0.957	0.036
BendTwin Ours	0.0079	0.0153	71.5	26.755	0.957	0.035

↓ Lower is better. ↑ Higher is better. CD: Chamfer distance; Track: tracking error. Values reproduced from Table 1 of the paper.

Dense → sparse

Fewer nodes, with bending constraints.

Matched-graph ablations test sampling ratios of 0.1 and 0.05. BendTwin improves on the axial-only baseline as connectivity is reduced. A separate experiment removes all interior nodes to examine a surface-only physical representation.

Read the ablation studies

03 Cite the work

Reference

BibTeX

@inproceedings{jing2026bendtwin,
  title={BendTwin: Robust Dense-to-Sparse Physical Reconstruction
         with Bending-Aware Differentiable Spring-Mass Models},
  author={Jing, Yixiong and Wang, Qi and Chen, Lin and Jiang, Junwei
          and Wang, Guangming and Wu, Haibing and Wysocki, Olaf
          and Ma, Wanli and Sheil, Brian},
  booktitle={NeurIPS 2026 Workshop (PUDM)},
  note={Oral presentation},
  year={2026},
  url={https://arxiv.org/abs/2608.06164}
}