OpenAI is directing robots on how to move using references created in virtual reality.
Another day, another exciting use for the versatile technology that is virtual reality. That’s what it feels like at least, considering nearly every day we’re waking up to fresh new possibilities and critical advances within the budding industry. We’ve seen augmented reality food menus, VR peripherals in the form of samurai swords & ping pong paddles, even an insane head-mounted display forcing you to aggressively poke your buddies in the face.
Now thanks to the creative minds over at OpenAI, VR has found yet another way to make its services useful: training robots to eventually conquer the Earth as our overlords, thus rendering all of humankind obsolete.
Okay so maybe I’m being just a little dramatic. OpenAI isn’t developing Skynet, but it’s still some relatively groundbreaking work in the field of robotics and simulation-based training that deserves some attention. Using their brand new algorithm, one-shot imitation learning, OpenAI has developed a new form of robotic communication that allows them to easily train a machine on certain functions & movements by demonstrating the desired action in virtual reality. However the coolest part has to be the fact that machines subjected to the simulations can perfectly replicate the actions after being shown them only once.
The Elon Musk-backed organization utilizes a system of dual neural networks that enable machines to learn by observing and decoding recorded actions. The ‘visual network’ takes an image from the robot’s camera and outputs a state indicating the positions of the objects. The second network, referred to as the ‘imitation network,’ observes a demonstration and interprets what is happening as well as how it is being done. By learning from thousands of demonstrations on various tasks the algorithm actually allows the machine to not only replicate scenarios, but adapt in order to accomplish goals in unpredictable circumstances as well. See?! Skynet!
“Nothing in our technique is specific to blocks,” says Josh Tobin, a researcher at OpenAI, in the video above. “This system is an early prototype, that will form the backbone of the general-purpose robotics systems we’re developing here at OpenAI.”
It appears as though the sky’s the limit for this open-ended VR system. While the actions and results of the robots may be small now, even simpler movements could eventually revolutionize the assembly lines of various industries and plants. Engineers and manufacturers could easily train massive amounts of robotic machinery capable of handling tedious aspects in the productions of mass-produced items such as vehicles or appliances. With the right demonstrations and the necessary advances in hardware, these intelligent machines even have the potential to handle more intricate and delicate functions which normally require a human touch.