Grasp Pose Learning from Human Demonstration with Task Constraints

To learn grasp constraints from human demonstrations, we propose a method that combines data-driven grasp constraint learning and one-shot human demonstration of tasks. By presenting task constraints in a GMM-based gripper-independent form, the task constraints are learned from simulated data with self-labeled grasp quality scores. By observing a human demonstration of the task and a real-world object, the learned task constraint model can be utilized to infer both the unknown grasping task and the probability density distributions of the task constraints on the object point cloud. In addition, we extend the superquadric-based grasp estimation method for reproducing the grasping task with 2-finger grippers. The task constraints restrict the searching scope of the grasp pose, so the geometrically best grasp pose within the task-constrained regions can be obtained. The effectiveness of our methodology is verified in experiments with a UR5 robot with a 2-finger gripper.