The document discusses a new approach for robots to learn grasping without human supervision. The robot collects its own training data by executing random grasps on objects and using force sensors to recognize successful grasps. As of now, the robot has executed over 17,000 random grasps, with around 1,300 successful grasps recorded along with image and depth data. This dataset is larger than existing human-labeled datasets and also includes information on failed grasps. The goal is for the robot to autonomously learn knowledge and understanding of grasping from its own experiences rather than relying on human knowledge transfers.