IPal Agent: Industrial Palletizing Embodied Agent With Interpretable Robotic Strategy Generation Across Diverse Objects and Patterns

In flexible manufacturing, industrial palletizing robots handle diverse objects and palletizing patterns. However, existing robots in rigid manufacturing mainly rely on predefined trajectories or rules, limiting their adaptability. Although embodied intelligence models for robotic manipulation have advanced rapidly, their closed-box decision-making behavior hinders deployment in industrial palletizing. Unlike pick-and-place, palletizing is a sequential long-horizon task that requires ordered grasp and precise place under geometric and structural constraints. To address this, we propose IPal Agent, an industrial palletizing embodied agent for interpretable strategy generation. IPal Agent integrates two key skills, Palletizing Grasper and Palletizing Placer, to achieve orderly and stable grasp as well as precise place of objects. We further construct a palletizing knowledge base and propose a mechanism to guide the agent’s palletizing strategy generation. Grounded in this knowledge, IPal Agent performs interpretable long-horizon palletizing across multiple objects and diverse palletizing patterns. We also establish a task-level benchmark and validate IPal Agent through both simulation and real-world experiments, achieving an average palletizing success rate of 81.7% in simulation and 72.0% in real-world experiments.