Reproduce “ChainSplat: A Physics-Inspired Screw-Theoretic Model for Learning Deformable Linear Object Dynamics from Multi-View RGB Videos” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Seungyeon Kim, Noémie Jaquier

CHAINSPLAT, developed at KTH Royal Institute of Technology, presents a framework for learning 3D geometry, appearance, kinematics, and dynamics of deformable linear objects (DLOs) directly from multi-view RGB videos using a physics-inspired screw-theoretic model and Gaussian splatting. This approach achieved superior dynamics prediction and rendering fidelity compared to baselines, and enabled efficient model-based manipulation without relying on auxiliary depth inputs.

Key insights

  • Representing DLOs as compact screw-theoretic open-chain systems provides a low-dimensional yet physically consistent model, which is crucial for efficient dynamics learning and real-time control applications.
  • The integration of link-aware 3D Gaussian splatting with kinematics-aware priors enables robust learning of DLO geometry and appearance directly from RGB videos, effectively bypassing the noise and complexity associated with depth sensors and multi-stage reconstruction pipelines.
  • A two-stage optimization strategy, separating kinematics-aware geometry recognition from dynamics identification, allows for stable and accurate estimation of parameters in complex deformable object systems, leading to a more robust digital twin.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.