Study hard enough, kids, and maybe one day you’ll become a professional battle robot. Boston Dynamics set the standard in the business a few years ago by having people wielding hockey sticks try to stop Spot the four-legged robot from opening a door. Previously, in 2015, the distant federal research agency Darpa organized a challenge in which it forced awkward humanoid robots to embarrass themselves on an obstacle course. way outside the league of machines. (I asked you once, dear readers, to stop laughing at them, but I’ve since changed my mind.) And now here it is: the makers of robot dog Jueying have taught him a fascinating way to fend off a human antagonizer who hits him or pushes him with a stick.
A team of researchers from the Chinese University of Zhejiang – where the Jueying material was also developed – and the University of Edinburgh did not teach the Jueying how to recover after an assault, so much so that they let the robot figure it out. It’s a radical change from the way a hardware developer like Boston Dynamics does it. teach a robot to move, using decades of human experience to hardcode, line by line, how a robot is supposed to react to stimuli like, uh, a person’s foot.
But there has to be a better way. Imagine, if you will, a football team. Midfielders, forwards, and goalkeepers all do things that are generally football-like, like 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 pitch who can grab the ball with their hands without getting yelled at.
In traditional robot training methods, you will have to meticulously code all of these specialized behaviors. For example, how should actuators – the motors that move the limbs of a robot – coordinate to make the machine work like a midfielder? “The reality is, if you want to send a robot out into the wild to do a wide variety of different tasks and missions, you need different skills, right?” says University of Edinburgh robot Zhibin Li, corresponding author on a recent article in the newspaper Scientific robotics describing the system.
Li and his colleagues began by training the 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 particular skill, such as trotting or standing up if it fell on its back. If the virtual robot tried something that brought it closer to the goal, it got a digital reward. If he did something non-ideal, he has a numerical demerit. This is called reinforcement learning. After several of these guided trial and error attempts, the simulated robot would become an expert in a skill.
Compare that to a robot’s traditional line-by-line method of coding to do something as simple as climbing stairs.this actuator turns as much, this other actuator turns as much. “The AI approach is very different in that it captures experience, that the robot has tried hundreds of thousands of times, even millions of times, ”Li explains.“ So in the simulated environment, I can create all possible scenarios. I can create different environments or different configurations. For example, the robot can start in a different pose like lying on the floor, standing, falling, etc. “