Revisiting Euler-Angle Regression with Kolmogorov-Arnold Networks

Forum topic · 小凯 · 2026-07-14

Summary

This arXiv paper (2607.09650) by Yangting Sun, Zijun Cui, and Yufei Zhang addresses the instability of Euler-angle regression in computer vision tasks. Euler angles are widely used to parameterize rotations in articulated robots, biomechanical models, and other systems, but their regression is unstable due to discontinuities and singularity (gimbal lock) issues. The authors introduce a new framework that combines range-aware Euler modeling with Kolmogorov-Arnold Networks (KAN), which replace fixed node activations with learnable univariate functions on edges. Their theoretical analysis shows that bounded Euler ranges induce a nearly additive regression structure, which aligns well with KAN's additive functional form. Experiments on controlled rotation regression, object pose estimation, and robotic and human inverse kinematics demonstrate consistent improvements in accuracy, convergence speed, and efficiency compared to existing approaches.

Paper Overview

Research Field: Computer Vision (CV)

Authors: Yangting Sun, Zijun Cui, Yufei Zhang

Publication Date: 2026-07-10

arXiv: 2607.09650

Abstract

In systems such as articulated robots and biomechanical models, rotations are parameterized by Euler angles, but their regression is unstable due to discontinuities and singularities. This paper introduces a new framework that combines range-aware Euler modeling with Kolmogorov-Arnold Networks (KAN), which replace fixed node activations with learnable univariate functions on edges.

Key Contributions

Range-aware Euler modeling: A formulation that constrains Euler angle ranges to improve regression stability.
KAN integration: Leverages Kolmogorov-Arnold Networks, which use learnable univariate functions on edges instead of fixed activation functions at nodes.
Theoretical analysis: The authors show that bounded Euler ranges induce a nearly additive regression structure — a property that aligns naturally with KAN's additive functional form.

Experimental Results

The framework was evaluated on:

Controlled rotation regression
Object pose estimation
Robotic and human inverse kinematics

Across all tasks, the method achieved consistent improvements in accuracy, convergence speed, and efficiency.

---

*Originally posted on zhichai.net forum. Auto-collected on 2026-07-14.*