基於強化學習結合KAN網路和注意力機制的創新混合架構應用於多機器人避障系統

Abstract

本研究提出一種結合 Kolmogorov-Arnold Network（KAN）與多層感知器(Multi-layer Perceptron , MLP）的創新架構，用於解決多機器人系統中的避障問題。隨著機器人技術的發展，多機器人系統在複雜環境中的運作日益普及，有效的避障策略成為確保系統安全與高效運作的關鍵。本研究設計了一種整合卷積神經網路（Convolutional Neural Network , CNN）、注意力機制（Attention）與 KAN 的混合架構，結合近端策略優化(Proximal Policy Optimization , PPO)算法進行強化學習訓練。實驗結果表明，與傳統的 CNN-MLP 架構相比，所提出的 CNN_ATT_MLP_KAN-PPO 架構在參數效率、學習效率和泛化能力方面均具有顯著優勢，特別適用於複雜環境和大規模多機器人系統。研究結果不僅驗證了 KAN 網路在實際應用中的價值，也為多機器人協作系統的發展提供了新的技術路徑。

This research proposes an innovative architecture combining Kolmogorov-Arnold Network (KAN) with Multi-layer Perceptron (MLP) to solve obstacle avoidance problems in multi-robot systems. As robotics technology advances, the operation of multi-robot systems in complex environments is becoming increasingly common, making effective obstacle avoidance strategies crucial for ensuring system safety and operational efficiency. This study designs a hybrid architecture integrating Convolutional Neural Networks (CNN), Attention mechanisms, and KAN, combined with the Proximal Policy Optimization (PPO) algorithm for reinforcement learning training. Experimental results demonstrate that the proposed CNN_ATT_MLP_KAN-PPO architecture shows significant advantages in parameter efficiency, learning efficiency, and generalization capability compared to traditional CNN-MLP architectures, particularly suitable for complex environments and large-scale multi-robot systems. The research findings not only validate the value of KAN networks in practical applications but also provide new technical pathways for the development of multi-robot collaborative systems.