← Back ICRA 2026

KAN Policy: Learning Efficient and Smooth Robotic Trajectories Via Kolmogorov-Arnold Networks

Zikang Chen, Fei Gao, Ziya Yu, and Peng Li

PDF

AI summary

Key figure (auto-extracted from paper)

Integrating Kolmogorov-Arnold Networks into Diffusion Policy significantly improves robotic trajectory smoothness, execution speed, and success rates in both simulation and real-world tasks.

Kolmogorov-Arnold Networks Diffusion Policy Robotic Control Smooth Trajectories Visuomotor Learning Continuous Representations

Problem

Current Diffusion Policy models rely on CNNs or Transformers with discrete feature processing, which fractures continuity and limits the generation of efficient, smooth motion trajectories for robotics.

Approach

The authors introduce Kolmogorov-Arnold Networks (KANs) into Diffusion Policy by designing an Embedding KAN (Emb-KAN) for CNN-based models and applying Group-KAN to Transformer-based models to enable continuous, spline-parameterized feature representations.

Key results

First integration of KANs into Diffusion Policy for robotics
Novel Emb-KAN preserves continuity in CNN-based latent spaces
Group-KAN adaptation enables continuous Transformer representations
Up to 53.8% higher success rate and 29.4% smoother motions in real-world tests

Why it matters

Provides a practical, architecture-level upgrade for visuomotor policy learning that enhances safety and efficiency for real-world robotic deployments.

Abstract

Modern robotic visuomotor policy learning has witnessed significant progress through Diffusion Policy (DP) frameworks built upon Convolutional Neural Networks (CNNs) and Transformers. Despite their empirical success, these archi- tectures remain fundamentally constrained by their relatively discrete computational nature, inherently limiting their capacity to generate efficient and smooth motion trajectories. To address this challenge, we introduce Kolmogorov-Arnold Networks (KANs) into Diffusion Policy learning. The proposed KAN Policy (KP) leverages KANs’ intrinsic continuity through learnable base- parameterized activation functions, thereby producing contin- uous trajectories with shorter execution time and fewer jerks. Specifically, we design a novel Embedding KAN (Emb-KAN) for CNN-based models, which preserves structural continuity in high-dimensional latent spaces through adaptive spline em- beddings. Besides, we apply Group-KAN to Transformer-based models for learning continuous representations. Across main simulation experiments, KP achieves average improvements of 6.06%, 8.03%, and 26.4% in terms of success rate, execution time, and smoothness, respectively. Similarly, in real-world ex- periments, KP achieves average improvements of 53.8%, 7.89%, and 29.4% across the same metrics.

Index terms

Deep Learning Methods Motion and Path Planning Learning from Demonstration