Optimal mobile robot routing with neural network and kernel-based dimensionality reduction in unknown environments

In this article, an advanced control system for mobile robots is presented, which combines nonlinear dimensionality reduction via kernel PCA with a soft-interval kernel and a dual neural network architecture (TFNN and WFNN) to enable optimal, fast, and safe navigation in unknown and noisy environments. The main innovation lies in designing an integrated framework that reduces high-dimensional sensory data from 1,067 to 18 key features—while retaining over 99% of variance—thus eliminating noise and improving data quality, and enabling targeted modeling of both target-following and boundary-following behaviors through two specialized neural networks. The system is validated using real-world data collected from a robot equipped with a SICK LMS200 sensor, resulting in a reduction of path error to 3.1 cm (a 62.79% improvement over the noisy baseline), a slip coefficient of 0.022, a zero collision rate, execution time reduced to 1.09 s, and path length minimized to 14.5 m. Furthermore, with variance coverage enhanced to 99.9%, model convergence time reduced from 25 to 10 epochs, and the interpretability index increased to 8.9; the proposed algorithm demonstrates superior efficiency, safety, and reliability when faced with real and noisy datasets. These achievements highlight the quantitative and qualitative superiority of the proposed method compared to leading approaches such as RRT*, PSO-Fuzzy, and DQN, establishing a new scientific foundation for the development of next-generation intelligent navigation systems.