Skeleton-Guided Rolling-Contact Kinematics for Arbitrary Point Clouds via Locally Controllable Parameterized Curve Fitting

Qingmeng Wen, Ze Ji, Yu-Kun Lai, Mikhail M. Svinin, Seyed Amir TafrishiIROS 2025 · DOI

Abstract

Rolling contact kinematics traditionally require continuous, differentiable ($C^2$) surface representations to calculate curvature and motion. However, real-world robotic sensors typically produce discrete, noisy point clouds. This paper introduces a framework that bridges this gap by using "curve skeletons" to define a reference frame and Fourier series to fit smooth surfaces to cross-sections. This approach allows robots to perform smooth rolling maneuvers on irregular objects without the computational overhead of full mesh reconstruction.

Introduction

In robotics, rolling contact is essential for tasks ranging from fingertip manipulation to spherical robot navigation. The mathematical foundation for this motion, formulated by Montana in the 1980s, relies on smooth surface geometry. Real-world robots, however, perceive the world through LiDAR or RGB-D sensors, resulting in point clouds that lack the necessary differentiability. Converting these into smooth surfaces is often computationally expensive, fails with incomplete data, or relies on oversimplified symmetry assumptions.

This research proposes a method to derive continuous rolling kinematics directly from raw point cloud data by transforming discrete points into differentiable manifolds through a skeleton-guided pipeline.

Methodology

The proposed framework transforms discrete point clouds into a continuous mathematical representation through a multi-step process:

  1. Skeletonization: A Laplacian-based contraction method collapses the 3D point cloud into a 1D "curve skeleton," representing the object's medial axis and providing a natural reference frame.
  2. Adaptive Slicing: The object is sliced into cross-sections perpendicular to the skeleton's tangent, reducing the 3D problem into a series of 2D shape-fitting tasks.
  3. Fourier Curve Fitting: Each 2D cross-section is modeled using a Fourier series. Because Fourier series are inherently smooth and differentiable, they provide the first and second derivatives required for rolling equations.
  4. Locally Parameterized Kinematics: A new set of kinematic equations is derived using a 2D manifold coordinate system ($u_g, v_g$), where $v_g$ represents the position along the skeleton and $u_g$ is the polar angle on the Fourier-fitted slice.
The multi-step pipeline for transforming point cloud inputs into contact position outputs.

By using the skeleton as an anchor, the system avoids the need for global symmetry and captures the natural "intent" of the shape. The Fourier fitting acts as a filter, smoothing out sensor noise while capturing the essential geometry of the object.

Example of the methodology applied to a carrot point cloud, showing the curve skeleton and the resulting Fourier fit in both polar and Cartesian coordinates.

Experimental Validation

The framework was validated using real-world experimental data from the OmniObject3D dataset, focusing on three objects with distinct geometric challenges: a carrot (tapered radius), a banana (highly curved skeleton), and an egg (asymmetric structure).

Key results from MATLAB simulations include:

  • Stable Trajectories: The system maintained smooth, physically plausible paths on all models, even when simulated slippage was introduced.
  • Differentiability: The Fourier-based model successfully provided the curvature and metric forms needed to solve rolling contact equations without singularities.
  • Density Sensitivity: A clear correlation was observed between point cloud density and model stability; denser clouds (like the banana model) resulted in fewer radius fluctuations and smoother motion.

Discussion and Future Work

While the method demonstrates robustness on irregular shapes, it currently assumes a single-path skeleton. Future work will focus on extending the slicing strategy to handle multi-branch geometries, such as trees or complex anatomical structures. Additionally, while the simulations utilized real-world data, the next phase involves testing these kinematics on physical robotic hardware and integrating the framework into a full motion planner to decide optimal trajectories for manipulation.

Summary updated

BibTeX
@inproceedings{wen2025skeleton,
  author = {Qingmeng Wen and Ze Ji and Yu-Kun Lai and Mikhail M. Svinin and Seyed Amir Tafrishi},
  title = {{Skeleton-Guided Rolling-Contact Kinematics for Arbitrary Point Clouds via Locally Controllable Parameterized Curve Fitting}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11246194},
}
Figure