Study hard enough, children, and maybe one day you will become a professional robot fighter. A few years ago, Boston Dynamics set the standard for the field by trying to stop people using hockey sticks from preventing the quadrupedal robot Spot from opening a door. Previously, the far-right federal research agency Darpa presented a challenge in 2015 in which it forced clumsy humanoid robots to embarrass themselves in an obstacle course. manner outside the machine league. (I once asked you, dear readers, to stop laughing at them, but have since changed your mind.) And now, look: the makers of the Jueying Robot Dog have learned a fascinating way to make a human to repel opponent who kicks it. over or press it with a stick.
A team of researchers from Zhejiang University in China – where Jueying’s hardware was also developed – and the University of Edinburgh file the Jueying how to recover after an assault, as much as they let the robot notice it. This is a dramatic departure from how a hardware developer like Boston Dynamics teaches a robot how to move, using decades of human experience to hard code, line by line, as a robot is supposed to respond to stimuli such as a person’s foot.
But there has to be a better way. Imagine, if you will, a football team. Midfielders, goal scorers and a goalkeeper generally do soccer-like things 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 field who can grab the ball with their hands without shouting.
In traditional methods of training robots, you need to carefully code all the specialized behaviors. For example, how should the actuators – motors that move the limbs of a robot – coordinate to make the machine run like a midfield? “The reality is that if you want to send a robot to nature to do a wide variety of different tasks and missions, you have to have different skills, right?” says the robotist Zhibin Li of the University of Edinburgh, corresponding author on a recent article in the magazine Science Robotics describe the system.
Li and his colleagues began training the software that would lead to a virtual version of the robot dog. They developed a learning architecture with eight algorithmic ‘experts’ that would help the dog deliver complex behaviors. For each of these, a deep neural network was used to train the robot model of the robot to achieve a particular skill, such as jogging or correcting should it fall on its back. If the virtual robot tries something that sets it closer to the goal, it gets a digital reward. If it did something non-ideal, it got a digital downfall. This is known as reinforcement learning. After many of these guided attempts at trial and error, the simulated robot would become an expert in a skill.
Compare that to the traditional way of coding a robot to do something as simple as climbing stairs—this drive turns so much, this other drive turns so much. ‘The AI approach is very different in the sense that it is captured experience, which the robot has tried hundreds of thousands of times or even millions of times, ”says Li. ‘So I can create all possible scenarios in the simulated environment. I can create different environments or different configurations. For example, the robot may start in a different position, such as lying on the ground, standing, falling, and so on. ”