A Grasping and Placement Learning Framework for Workpieces Packing in Complex Scenes

Online 3D workpiece packing is crucial in logistics, but achieving high space utilization without prior knowledge of incoming item sequences remains challenging. Current methods struggle to capture dynamic 3D spatial topologies and heavily rely on rigid heuristics, which limits exploration diversity and causes space fragmentation. A deep reinforcement learning framework integrating a spatially weighted graph attention network and a diffusion model is proposed. The network accurately extracts the global topological dependencies and local geometric constraints of the packing state. Concurrently, the diffusion model synthesizes highly diverse, physically feasible candidate placement poses, fundamentally expanding the agent’s exploration boundary. Experiments demonstrate that our method significantly outperforms baselines in space utilization and efficiency. Its effectiveness and robustness are further validated in a real robotic system, offering a practical hybrid reinforcement learning solution for online 3D workpiece packing.