Semi-Supervised Voronoi-TSVM Path Planning for Automated Guided Vehicles under Multi-Modal Perception

Abstract

This work proposes a fusion optimization method combining Voronoi diagram and Transductive Support Vector Machine (TSVM) to address the path planning problem of Automated Guided Vehicle (AGV) in complex environments. First, this method uses a Voronoi diagram to construct the spatial skeleton of the environment and generate an initial path. Second, it introduces TSVM to perform semi-supervised classification and optimized screening on path segments, to screen out safer and smoother segments. Finally, this method ensures the safety of the path while improving its smoothness and operational efficiency. In evaluation indicators, path smoothness is defined as the standard deviation of path angle changes (unit: radian), and a smaller value indicates a smoother path; collision rate refers to the proportion of path points where the minimum distance from obstacles is lower than the safety threshold. The experimental results show that the optimized model has a path length of 42.987 meters (m) under LiDAR data; it is remarkably shorter than that of Convolutional Neural Network - Long Short-Term Memory (CNN-LSTM) (46.879m) and Deep Reinforcement Learning - Model Predictive Control (DRL-MPC) (45.212m). This indicates that the optimized model can offer more efficient path choices. Concerning computation time, this model performs excellently, with 1.345 seconds (s) under LiDAR data, much lower than CNN-LSTM's 2.346s and DRL-MPC's 1.879s, demonstrating its high efficiency in path planning. Regarding path smoothness, the optimized model achieves 0.208, superior to CNN-LSTM's 0.345 and DRL-MPC's 0.298, thus reducing vibrations and path deviations. Moreover, this model has a collision rate of 0.012 under LiDAR data, remarkably lower than CNN-LSTM's 0.062 and DRL-MPC's 0.045. Consequently, this work develops a novel path planning strategy integrating classification learning with spatial mapping. It also validates the strategy's universality and robustness in multi-source perception scenarios, providing theoretically sound and engineering-feasible optimization solutions for intelligent AGV systems.