DeformX

A Versatile Co-Simulation Framework
for Deformable Linear Objects

Yi Yang^1,3,4,*, Xiang Fei^1,*, Lehong Wang^1,*, Chenhao Li², Zilin Dai⁵, Henry Kou¹, Howie Choset¹, Lu Li¹

¹The Robotics Institute, Carnegie Mellon University   ²Dept. of Mechanical Engineering, Carnegie Mellon University

³School of Ocean and Civil Engineering, Shanghai Jiao Tong University   ⁴Zhiyuan College, Shanghai Jiao Tong University   ⁵Harvard University

* Equal contribution

Simulating deformable linear objects (DLOs) such as wires, cables, and ropes with both visual realism and physical accuracy remains a significant challenge. We present DeformX, a co-simulation framework integrating a Cosserat rod physics engine with NVIDIA Isaac Sim. DeformX simulates dynamics, self-collisions, and interactions with free-form meshes, while using mesh skinning to map rod deformations to CAD models for high-fidelity visualization. It is the first framework combining realistic visualization, principled physics, and robot learning compatibility.

Key Contributions

Co-Simulation Framework

Integrates a Cosserat rod physics engine with NVIDIA Isaac Sim via a multi-rate coupling scheme, bridging time scales of 10^-5s (DLO physics) and 10^-2s (rendering/robot control). A modular Python interface enables scene authoring with support for dynamics, self-collisions, and interactions with free-form meshes.

Mesh Skinning
Maps discrete Cosserat rod deformations onto CAD models for high-fidelity visualization.

Free-Form Mesh Contact
Supports DLO interactions with arbitrary meshes using BVH-accelerated collision detection.

WireSeg-32k Dataset
32,000 synthetic wire segmentation images with instance masks across Easy/Medium/Hard tiers.

WireSeg-32k Dataset & Segmentation Results

32,000 synthetic images from 300+ simulation runs across Easy/Medium/Hard tiers in wire-on-plane, flying wires, and data center scenarios.

Fine-tuning SAM3 on DeformX-generated data yields 10.2% mAP@75 improvement in real-image wire instance segmentation.

Sim-to-Real Robot Learning

Physics validation: robot-driven rope motion with motion capture comparison.

Goal-conditioned dynamic rope manipulation: simulation vs. real-world on UR5e.

A rope-swinging hit-target policy trained entirely in DeformX achieves 6.6 cm mean target-hitting error when deployed on a real UR5e robot, demonstrating effective sim-to-real transfer.

Paper

Submitted to IROS 2025.

Acknowledgements

This work is supported by The Robotics Institute at Carnegie Mellon University.

Contact

If you have any questions, please feel free to contact Lehong Wang.