This study addresses the problem localization deviation caused by cumulative wheel odometry errors in Automated Guided Vehicles (AGVs) operating in complex environments by proposing an adaptive localization method based on multi-sensor fusion. Within an Extended Kalman Filter (EKF) framework, the proposed approach integrates internal sensor predictions with external positioning data corrections, employing an adaptive weighting algorithm to dynamically adjust the contributions of different sensors. This effectively suppresses errors induced by factors such as ground friction and uneven terrain. The experimental results demonstrate that the method achieves a localization accuracy of 13 mm, and the simulation results show a higher accuracy of 10 mm under idealized conditions. The minor discrepancy is attributed to unmodeled noise and systematic errors in the complex real-world environment, thus validating the robustness of the proposed approach while maintaining robustness against challenges such as Non-Line-of-Sight (NLOS) obstructions and low-light conditions. The synergistic combination of LiDAR and odometry not only ensures data accuracy but also enhances system stability, providing a reliable navigation solution for AGVs in industrial settings.

To address the challenges of low efficiency, instability, and difficulties in meeting multiple constraints simultaneously in multi-AGV (Automated Guided Vehicle) task scheduling for intelligent manufacturing and logistics, this paper introduces a scheduling method based on multi-feature constraints and an improved deep reinforcement learning (DRL) approach (Improved Proximal Policy Optimization, IPPO). The method integrates multiple constraints, including minimizing task completion time, reducing penalty levels, and minimizing scheduling time deviation, into the scheduling optimization process. Building on the conventional PPO algorithm, several enhancements are introduced: a dynamic penalty mechanism is implemented to adaptively adjust constraint weights, a structured reward function is designed to boost learning efficiency, and sampling bias correction is combined with global state awareness to improve training stability and global coordination. Simulation experiments demonstrate that, after 10,000 iterations, the minimum task completion time drops from 98.2 s to 30 s, the penalty level decreases from 130 to 82, and scheduling time deviation reduces from 12 s to 0.5 s, representing improvements of 69.4%, 37%, and 95.8% in the same scenario, respectively. Compared to genetic algorithms (GAs) and rule-based scheduling methods, the IPPO approach demonstrates significant advantages in average task completion time, total system makespan, and overall throughput, along with faster convergence and better stability. These findings demonstrate that the proposed methodology enables effective multi-objective collaborative optimization and efficient task scheduling within complex dynamic environments, holding significant value for intelligent manufacturing and logistics systems.