To address the high sensing cost, uneven material distribution, and safety–efficiency trade-off in close-range wheel loader–dump truck collaborative unloading, this study proposes a perception–task–control framework for autonomous unloading. A complementary front–rear vision configuration is used to perceive the dump-truck bed under varying relative viewpoints, and the estimated bed pose is further transformed into executable unloading targets. To improve load distribution, a partition-aware task-generation strategy is developed, by which the unloading objective is extended from a single target point to sequential zone-level targets. An event-triggered two-stage reinforcement learning controller is then designed to organize the unloading process. The first stage guides the loader toward a perception-enabled region, while the second stage performs vision-guided precision alignment and coordinated lifting according to the current zone-level target. A closed-loop co-simulation environment is constructed using MATLAB/Simscape R2025b and Unreal Engine, and field-test data are used for simulation–field response comparison. The simulation results under representative operating conditions show that the proposed framework can complete sequential zone-level unloading without collision under the tested conditions. The quantitative results support the effectiveness of the method in terms of target completion, completion time, terminal positioning accuracy, lifting completion, and collision avoidance. The field-test comparison further indicates that the developed simulation model can reproduce the main trajectory, articulation-angle, and lifting-cylinder displacement responses of the wheel loader during unloading. These results demonstrate the feasibility of integrating low-cost visual perception, partition-aware task generation, and two-stage learning-based control for autonomous wheel-loader unloading.
Liu et al. (Tue,) studied this question.