The paper addresses the challenging problem of grasping novel objects using vision, particularly those that have never been seen before. The authors propose a learning algorithm that does not require a 3D model of the object and instead identifies grasp points directly from images. This approach is contrasted with dense stereo methods, which attempt to triangulate every point in an image, often failing to provide accurate 3D models. The algorithm uses supervised learning with synthetic images to train a model that predicts grasp points in real images. The algorithm is tested on two robotic manipulation platforms and successfully grasps a wide variety of objects, including plates, tape rolls, jugs, cellphones, keys, screwdrivers, staplers, thick coils of wire, and a power horn. The paper also applies the method to the task of unloading items from dishwashers, achieving an average grasp success rate of 80.0% for plates, bowls, mugs, and wine glasses. The authors discuss the limitations of their approach, such as the difficulty in grasping objects with narrow trajectories and the need for more advanced robotic arms with haptic feedback.The paper addresses the challenging problem of grasping novel objects using vision, particularly those that have never been seen before. The authors propose a learning algorithm that does not require a 3D model of the object and instead identifies grasp points directly from images. This approach is contrasted with dense stereo methods, which attempt to triangulate every point in an image, often failing to provide accurate 3D models. The algorithm uses supervised learning with synthetic images to train a model that predicts grasp points in real images. The algorithm is tested on two robotic manipulation platforms and successfully grasps a wide variety of objects, including plates, tape rolls, jugs, cellphones, keys, screwdrivers, staplers, thick coils of wire, and a power horn. The paper also applies the method to the task of unloading items from dishwashers, achieving an average grasp success rate of 80.0% for plates, bowls, mugs, and wine glasses. The authors discuss the limitations of their approach, such as the difficulty in grasping objects with narrow trajectories and the need for more advanced robotic arms with haptic feedback.