Grasp Planning for a Hybrid Compliant Robotic Gripper

Kurzfassung

Bin-picking for warehouse automation is still an open problem, despite the major advancements of the recent years in robotic grasping and manipulation. A successful automated bin-picking system has to be able to continuously detect the objects present in the bin, plan a grasping approach and execute it. In the case of hybrid grippers, which offer more than one grasping modality, the problem is complicated by the necessity of choosing between the available modes, or choosing how to combine them to have an effective, stable and robust grasp. This work’s contribution lies in the introduction of the first grasp planner for the Hybrid Compliant Gripper (HCG) developed at the German Aerospace Center (DLR). The gripper has two fingers that can move independently, with tendons of variable stiffness, and it is mounted on a robotic arm. It has two grasping modes: fingers and suction. As a first step towards a complete grasp planner, the approach presented here considers single unknown objects placed within the workspace of the robot. The objects are assumed to be rigid and non-transparent. The grasp planner feeds point clouds extracted from RGB-D images of the object to a deep neural network that outputs the semantic segmentation of the input point cloud, i.e., for each point in the point cloud, the network outputs two binary labels, which indicate if the point is suitable for a grasp with fingers or not and if it is suitable for a suction grasp or not. The network architecture is based on PointNet++ modules and it has three heads: One that learns to reconstruct the normal vector to the surface in each point, acting as an autoencoder; One that outputs the semantic segmentation for grasps with fingers, and one that outputs the semantic segmentation for suction grasps. The network was trained on a custom generated synthetic grasp dataset, based on the novel EGAD object dataset, a procedurally generated dataset of objects specifically designed to expand the graspability space and study grasping and manipulation problems. During the data generation procedure, each object in the EGAD dataset is loaded in a PyBullet simulation and used to generate point clouds Each point of each point cloud is then annotated with a quality metric for grasps with fingers and a quality metric for a suction grasp. The capability of the network to generalize to unseen objects belonging to a different dataset, the well-established YCB dataset, was verified. First, laser scans of the objects were fed to the data generation procedure to obtain ground truth labels; Then, these were compared to the labels output by the network for the same point clouds extracted from the laser scans. Finally, the ability of the presented grasp planner to plan executable grasps in real-world scenarios is demonstrated by using the HCG to grasp a selection of YCB objects