Real-time six degrees of freedom grasping of deformable poultry legs in cluttered bins using deep learning and geometric feature extraction

Abstract

Grasping deformable natural food products in pick-and-place applications presents a challenging problem in robotics, particularly in cluttered environments. Traditional techniques, such as centroid-based grasping and grasping along the principal axis, often fail when applied to deformable and complex-shaped objects. A real-time grasp pose estimation algorithm was proposed in this paper for robotic pick-and-place tasks involving deformable objects. Instance segmentation, keypoint detection, and stable six degrees of freedom (6-DoF) grasp pose estimation were jointly performed within a unified framework, while avoiding collisions using deep learning and geometric features. Grasp candidates on the objects were generated based on maximum curvature and convex hull methods, which were then filtered to identify the most stable grasp. By projecting 2D keypoints into the depth image, the grasping points were obtained in 3D. The keypoints were used to estimate the 3D orientation of the grasp. The success rate of grasp pose estimation was found to be 83.3 %, which increased to 93.7 % when object segmentation failures were removed. The algorithm was evaluated on a real robotic platform, demonstrating its ability to accurately grasp real poultry legs from a pile, with success rates of 93.8 % for simple scenes and 75.0 % for cluttered scenes. These results highlight state-of-the-art performance and showcase the efficacy and robustness of the proposed system.