Perception-constrained joint optimization of 3D railcar loading and executable loading paths via deep reinforcement learning

To address the challenges of complex 3D spatial states, strong dependence on cargo placement order, and difficulty in guaranteeing path executability during loading, this paper proposes a joint 3D modeling and executable loading-path optimization method for robotic loading in railway freight wagon scenarios. The method first acquires the box dimensions, pose, and loading area state based on 3D vision, and constructs a 3D state representation integrating height maps, occupancy masks, candidate placement points, and path reachability information. Subsequently, placed boxes, boxes to be placed, and virtual nodes are uniformly modeled as a node sequence, and a node mask mechanism is introduced to filter out boundary crossings, collisions, insufficient support, and unreachable paths. Furthermore, a Transformer policy network integrating 3D position encoding is designed to achieve joint decision-making for box selection, placement posture, and candidate positions. To verify the effectiveness of the method, experiments are conducted under different loading sequences, number of boxes, and module ablation conditions, and simulations are performed using a robotic platform. Experimental results show that the proposed method outperforms the comparative model in terms of space utilization, loading success rate, path cost reduction, and stability gain, especially demonstrating better robustness and executability in high-density loading scenarios.