Yongchun Liu

Autonomous loading operations in underground mining present complex engineering challenges due to the unpredictable interactions between load–haul–dump (LHD) machines and fragmented ore, as well as the dynamic and heterogeneous nature of muck piles. To address these challenges, this study presents a simulation-based deep reinforcement learning (DRL) autonomous loading controller tailored for LHD performing drawpoint mucking in block and panel caving and sublevel caving mines. The proposed method leverages the proximal policy optimization algorithm, trained within a high-fidelity simulation environment developed using the MuJoCo physics engine and the Gymnasium interface, and enhanced by domain randomization techniques to improve robustness across diverse loading scenarios. A novel reward function incorporating thrust, loading volume (LV), cycle-time-related efficiency, and action penalties, together with an optimized observation space integrating key sensor data, is designed to enhance learning effectiveness. Comprehensive simulation experiments show that the proposed controller outperforms a finite-state-machine-based controller, skilled human operators, and standard model-free RL baselines (twin delayed deep deterministic policy gradient, soft actor–critic, deep deterministic policy gradient) in terms of LV and volumetric loading rate, while maintaining a high success rate of loading under the tested conditions. Ablation studies further validate the critical role of individual reward components and observation parameters in achieving stable and efficient loading behavior. In addition, we explicitly model sensor errors (measurement noise and drift) for all onboard sensors and evaluate the robustness of the learned controller under moderate and severe sensor degradation and single-sensor failures, quantifying their impact on loading performance. This simulation-based study thus indicates that DRL is a promising and effective approach for autonomous LHD loading, with the potential to enhance operational efficiency and safety in underground mining environments as the controller is progressively transferred from simulation to real machines.

As a critical enabler of trackless mining operations, load-haul-dump (LHD) vehicles rely heavily on efficient and safe path planning to support the industrial deployment of autonomous driving systems. However, existing planning algorithms often suffer from heavy computational demands, limited adaptability to the constrained kinematics of articulated LHDs, and insufficient robustness in complex underground environments. To address these challenges, we propose a Hybrid A* path planning method guided by high-level reference states. The method first integrates the A* algorithm with generalized Voronoi diagrams to rapidly generate a reference state set along tunnel centerlines. This reference serves as a structural guide for downstream search. Then, a region-aware dynamic node expansion strategy and a multi-objective heuristic cost function are introduced to improve planning efficiency and path quality. Comparative evaluations across three representative underground tunnel scenarios demonstrate that the proposed approach outperforms baseline methods in terms of safety, efficiency, and robustness, providing a viable solution for motion planning of LHDs in confined underground spaces.

This study addresses collision detection in the unmanned loading of ore from load-haul-dump (LHD) machines into mining trucks in underground metal mines. Such environments present challenges like heavy dust, confined spaces, sensor occlusions, and poor lighting. This work identifies two primary collision risks and proposes corresponding detection strategies. First, for collisions between the bucket and tunnel walls, LiDAR is used to collect 3D point cloud data. The point cloud is processed through filtering, downsampling, clustering, and segmentation to isolate the bucket and tunnel wall. A KD-tree algorithm is then used to compute distances to assess collision risk. Second, for collisions between the bucket and the mining truck, a kinematic model of the LHD’s working device is established using the Denavit–Hartenberg (DH) method. Combined with inclination sensor data and geometric parameters, a formula is derived to calculate the pose of the bucket’s tip. Key points from the bucket and truck are then extracted to perform collision detection using the oriented bounding box (OBB) and the separating axis theorem (SAT). Simulation results confirm that the derived pose estimation formula yields a maximum error of 0.0252 m, and both collision detection algorithms demonstrate robust performance.