OpenAI trained a robot arm to grasp real objects using only crude, randomized simulated images. The system never saw real photos during training, yet transferred directly to physical robots by varying shapes, textures, lighting, and noise across thousands of simulated examples.
Dear fellow scholars, this is two minute papers with in this series we talk a lot about different AI algorithms that solve a variety of super difficult tasks. These are typically tested within a software environment in the form of a simulation program. However, this often leaves the question open whether these algorithms would really work in real world environments. So what about that? This work from OpenAI goes by the name domain Randomization and is about training an AI on relatively crude computer simulations in a way that can be transferred to the real world. The problem used to demonstrate this was localizing and grasping objects. Note that this algorithm has never seen any real images and was trained using simulated data. It only played a computer game, if you will. Now the question we immediately think about is what the term domain randomization has to do with transferring simulation knowledge into reality. The key observation is that using simulated training data is okay, but we have to make sure that the AI is exposed to a diverse enough set of circumstances to obtain knowledge that generalizes properly. Hence the term domain randomization. In these experiments the following parameters were heavily number of shapes and distractor objects on the table, positions and textures on the objects, table and the environment, number of lights, material properties and the algorithm was even exposed to some random noise as well in the images. And it turns out that if we do this properly, leaning on the knowledge of only a few thousand images, when the algorithm is uploaded to a real robot arm, it becomes capable of grasping the correct prescribed objects. In this case, the objective was span detection. Very amusing. I think the very interesting part is that it is not even using photorealistic rendering and light simulations. These programs are able to create high quality images that resemble the real world around us, and it is mostly clear that those would be useful to train such an algorithm. However, this only uses extremely crude data and the knowledge of the AI still generalizes to the real world. How about that? Thanks for watching and for your generous support and I'll see you next time. Sam.
