VC-SLAM: Optimizing Cloud–Edge VSLAM Transmission Based on Variable-Order Chebyshev-KAN

Abstract

In autonomous robot tasks, the cloud–edge collaborative visual simultaneous localization and mapping (VSLAM) effectively addresses the issue of computing resource limitations in mobile robots. To improve tolerance for transmission delay and enhance real-time performance, it is crucial to reduce the transmission data volume effectively. With the development of neural implicit VSLAM, network model compression technology helps reduce transmission data volume. The Chebyshev-Kolmogorov–Arnold network (KAN) model is a potential solution with a small network size, while its low training efficiency cannot meet the real-time requirement. In this article, we design a variable-order Chebyshev-KAN network (VC-KAN), utilizing the order variation properties of Chebyshev polynomials. The VC-KAN model optimizes network training for implicit representation, thereby improving real-time performance while maintaining the advantage of small transmission data volume. A cloud–edge collaborative VSLAM system, VC-SLAM, is also developed based on the proposed VC-KAN, achieving efficient collaboration. Experimental results indicate that compared to the existing methods, we reduce the amount of transmitted data by 52.00% and improve accuracy by 38.69%. We shorten the average training time of the implicit representation network by 82.70%, improving the real-time performance for cloud–edge collaborative VSLAM.