Depalletisation humanoid torso: Real-time cardboard package detection based on deep learning and pose estimation algorithm

This paper presents a system for efficiently depalletising textureless and versatile cardboard boxes, even when they are tightly packed together, by utilizing deep learning and point clouds. The system uses an industrial humanoid torso equipped with a 6 DoF dual-arm and an RGB-D camera allowing for fast, accurate, and adaptable object handling without the need for additional sensors or setup. Operating in a warehouse environment, the robot is responsible for loading/unloading pallets of cardboard packages and placing them onto conveyor belts for further processing. The system uses ROS interface for communication, control, and MoveIt for dual-arm path planning and achieves impressive F1 scores of 0.81 and 0.90 for single-face and multi-face boxes, respectively, as demonstrated through real-time testing. These results provide valuable insights into the system’s capabilities and potential future improvements.