Multi-class 4-DoF carton box detection for heterogeneous robotic container unloading

Container unloading is one of the most challenging tasks in depalletizing applications and has not been well addressed. Because it is non-trivial to accurately detect irregularly packed boxes and efficiently unload boxes in confined spaces, in this paper, we investigate both the detection step and the motion planning step in the container unloading task. First, we introduce the multi-class 4-DoF box detection into container unloading tasks and build a detection method based on YOLOv5. To generate the unloading sequence of boxes, a heuristic rule is proposed by using human prior knowledge. And to match the categorical attributes of different types of boxes and task setups, we design suitable motion trajectories. Based on the proposed methods, a complete robotic unloading system is developed. We evaluate our system via real-world experiments and the results indicate the efficiency and reliability of our robotic system. The video demonstration can be found at: https://youtu.be/nYvCbNoYi5Y.