A robot dog has been taught how to defend itself if it is antagonised by a human.
The mechanical animal, called Jueying, was developed at China’s Zhejiang University, and through a collaboration with researchers from the University of Edinburgh let the mechanical canine learn what to do when it was assaulted.
Roboticists used a deep neural network – an algorithm that is designed to mimic the human brain – to train a computer model of the robot dog. This gave it the ability to trot, or stand up again if it fell over.
If the robot did something beneficial to its task, it was rewarded, and if it did not it was digitally noted as a negative behaviour – essentially a more sophisticated trial-and-error method that researchers use to train robots to learn.
“The AI approach is very different in the sense that it captures experience, which the robot has tried hundreds of thousands of times, or even millions of times,” University of Edinburgh roboticist Zhibin Li, corresponding author on a paper in Science Robotics, told Wired.
“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, such as lying down on the ground, standing, falling over, and so on.”
.@InfAtEd, @ZJU_China researchers devised a new deep learning architecture that can combine pre-learned skills to create new skills on the fly. A four-legged robot using the #AI quickly navigated complicated situations, such as recovering from a fall: https://t.co/ehpK3h8aYp pic.twitter.com/8KXsHdnFcj
— Science Robotics (@SciRobotics) December 11, 2020
Eight simulated ‘experts’ were trained in order to develop the robot’s various skills, such as walking, maintaining balance, running, and more. When the robot faced a difficult task, such as falling down over uneven surfaces, each of the algorithmic experts could spark into life and tell the machine what to do in order to get back up.
INDY/LIFE Newsletter
Be inspired with the latest lifestyle trends every week
Please enter your email addressPlease enter a valid email addressPlease enter a valid email address SIGN UPThanks for signing up to the INDY/LIFE newsletter{{#verifyErrors}} {{message}} {{/verifyErrors}} {{^verifyErrors}} {{message}} {{/verifyErrors}}The Independent would like to keep you informed about offers, events and updates by email, please tick the box if you would like to be contacted
Read our full mailing list consent terms here
INDY/LIFE Newsletter
Be inspired with the latest lifestyle trends every week
SIGN UPThe Independent would like to keep you informed about offers, events and updates by email, please tick the box if you would like to be contacted
Read our full mailing list consent terms here
As such, when a human comes over to the dog and pushes it to the floor, the robot knows how to return to its feet.
The end-goal of this research is to teach robots to learn in the same way that human babies do; rather than having to code specific actions and resistances for climbing stairs, for example, the robot could simply be ‘told’ that moving their limbs is enough, and then artificial intelligence could – eventually – achieve the correct movement.
A new reinforcement learning architecture designed by @InfAtEd, @ZJU_China could quickly calculate new ways to move and smoothly transition between different motions as it faced new scenarios it had never encountered during training: https://t.co/ehpK3h8aYp #AI #robot pic.twitter.com/JIfx2G7ipY
— Science Robotics (@SciRobotics) December 14, 2020
When robots enter everyday life, or are used in certain commercial applications, this way of learning will be vital, because there are so many uncontrollable variables that researchers cannot predict that it would be impossible to try and code a robot for every eventuality.
“A grand picture or grand vision,” Li said, “is that we will have more intelligent machines, which are able to combine flexible and adaptive skills on the fly, to handle a variety of different tasks that they have never seen before.”
