AsiaIndustrial NetNews: The human brain is powerful, but it has a serious limitation: You can’t download new information in real-time like The Matrix. However,robotsure.
Imagine a future where these Robots are connected to each other in the cloud, and when one of them learns something, they all learn. Let’s hope everything is good, like giving each other a hug. But the problem is, you can’t just teach a small car to learn to grasp, and then expect to give the same skills to a hulking bipedal robot.
But today, a new study from MIT’s Computer Science and artificial intelligence Laboratory takes a big step forward in enabling this seamless knowledge transfer. It all started with a little robot named Optimus and his friend Atlas, the famous 6-foot-tall humanoid robot. Researchers have taught Optimus, a bomb disposal robot, how to pull one tube from another.
First, they store some information about the different operations required for different objects. Then, they made it hold a sim card in its hand.
Claudia Perez-Dapino, a robotics expert and one of the authors of the study, said: “Imagine a 3D world in which a robot lives, and with a mouse, you can control its hands and move them. This way, the robot can be commanded without a professional programmer. It’s more intuitive for the operator because it’s similar to how humans learn: Toddlers acquire a skill, such as how to grab a dolls, and when they encounter new things, they can readjust their original skills.
Now, how do you transfer a robot’s skills to a robot many times its size? After all, the robot faces a new challenge: not falling face down. “Mathematically, it can be written as another set of constraints,” Perez-da Appino said, “if you can imagine it, put your weight on a certain area.” Essentially Say, the operator has to give this new robot some rules, like how to properly balance and how to perform the same tasks as the Optimus. Combine these rules with Optimus’ skills in handling pipes, and you’ll have a smooth ride. Of course, this is not an automatic conversion, this is just the beginning.
At present, Atlas can only complete the handover work in the simulator. But this development gives us a glimpse into a future where more robots can interact with humans. For example, they might teach themselves to take a pipe out of a pipe through a process called “reinforcement learning,” which basically keeps trying until they finally find the right one.
Imagine, in a factory setting, if a robot learns how to maneuver something more efficiently, it can spread the skills to its peers via the cloud. And, as Perez Dappino has shown, these skills can even allow various robots to work together.
Soon, robots will be able to think without human help and freely spread these skills, skills like hugging.
The Links: 3HAC032586-001 3HNA009885-001 IGBT