Recognizing the object is only the beginning
Image classification and robotic grasping ask different things of vision. QT-Opt makes the action-and-feedback part explicit.
A machine that names an object has answered a different question from one that must pick it up. The ImageNet paper trains a convolutional neural network to classify images into a thousand categories. Its reported competition results concern classification error. The architecture combines convolutional and fully connected layers; the authors describe GPU implementation and measures to reduce overfitting. That is a substantial perception result. It is not, by itself, a demonstration of a robot choosing and adjusting a grasp. [1]
QT-Opt's authors study vision-based robotic grasping through reinforcement learning. Their abstract describes closed-loop control: the robot updates its grasp strategy using its most recent observations, rather than selecting a grasp point and simply executing it. They report learning from more than 580k real-world grasp attempts and achieving 96% grasp success on unseen objects. They also describe regrasping, probing and repositioning objects, and responding to disturbances, using RGB images from an over-the-shoulder camera. These are the authors' experimental claims about grasping. The abstract supplies no basis for turning that result into a promise about every object, workplace or assignment. [2]
UAF's point is that better recognition and better control deserve separate questions. What does the system identify? What action does it choose? What can a new observation change while the action is underway? The two papers make that distinction concrete, but their classification and grasping scores are not interchangeable measures of intelligence. This comparison also establishes no claim that QT-Opt simply inherits the ImageNet network. Before promoting a perception result into a claim about autonomy, ask to see the action and its feedback. I would enjoy being promoted on the strength of a good look. You should probably require more. [1] [2]
