Study hard enough, children and maybe one day you will grow up and become a professional robot fighter. A few years ago, Boston Dynamics set the standard for the field by the fact that people handling hockey sticks try to prevent Spot the four-legged robot from opening a door. Earlier in 2015, the remote federal research agency Darpa hosted a challenge in which it forced clumsy humanoid robots to be embarrassed on an obstacle course. way outside the car league. (I once asked you, dear readers, not to laugh at them, but I have changed my mind since then.) And now, here: The producers of the robot dog Jueying have taught him a fascinating way to defend a human antagonist who hit him or push him with a stick.
A team of researchers from Zhejiang University in China – where Jueying hardware was also developed – and the University of Edinburgh did not teach Jueying how to recover from an assault, as much as they let the robot realize. It’s a dramatic departure from the way a hardware developer, such as Boston Dynamics, teaches a robot how to move, using decades of human experience in hard code, line by line, how a robot is supposed to move. to react to stimuli such as … the foot.
But there must be a better way. Imagine, if you want, a football team. Midfielders, strikers and a goalkeeper generally do essential things for football, such as running and kicking, but each position has its own specialized skills that make it unique. The goalkeeper, for example, is the only person on the field who can grab the ball with his hands without screaming.
In traditional robot training methods, you will need to meticulously codify all of these specialized behaviors. For example, how should actuators – motors that move the limbs of a robot – coordinate to make the machine work as a middleman? “The reality is, if you want to send a robot into the wild to do a wide range of different tasks and missions, you need different skills, right?” says University of Edinburgh robotist Zhibin Li, correspondent in a recent journal Robotic science describing the system.
Li and his colleagues began by training software that would guide a virtual version of the robot dog. They developed a learning architecture with eight algorithmic “experts” that would help the dog to produce complex behaviors. For each of these, a deep neural network was used to train the computer model of the robot to achieve a certain ability, such as trotting or straightening itself, if it fell on its back. If the virtual robot tried something that brought it closer to the target, it received a digital reward. If he did something non-ideal, he got a digital demerit. This is known as reinforcement learning. After many of these trial and error guided trials, the simulated robot would become an expert in a skill.
Compare this to the traditional line-by-line coding mode of a robot to do something as simple as climbing stairs …this servomotor transforms so much, this other servomotor transforms so much. “The AI approach is very different in the sense that it surprises the experience, which the robot has tried hundreds of thousands of times, or even millions of times, ”says Li. “So, in the simulated environment, I can create all possible scenarios. They can create different environments or different configurations. For example, the robot may start in a different position, such as lying on the floor, standing, falling, and so on. “