Watching someone kick a robot dog is one of those weirdly unsettling things, despite the fact that we know that the canine in question is just a collection of servos and other high tech components. However, it’s an important reaction to test, since it’s the kind of unexpected collision that a robot might need to deal with if it’s going to function in the real world: especially if it will be working in unsafe environments or travelling on unstable surfaces.
Fortunately, researchers from the Swiss Federal Institute of Technology (ETH) Zurich in Switzerland are willing to do this job so you don’t have to. And the results are already looking promising. Roboticists at the research institution have demonstrated how their ANYmal four-legged robot is capable of taking a kicking and keeping on ticking — or, well, keeping on walking at least. More impressively, this ability to recover from potential knockout blows doesn’t require any additional hardware, but instead the implementation of a new algorithm. Oh, and it requires way less physical kicking to test than previous attempts.
“The main contribution of [our latest research paper] is to demonstrate that such complicated behaviors can be trained using only simulated data,” Jemin Hwangbo, the researcher who led the study, told Digital Trends. “Previously, simulation was not accurate enough to train performant control policies. Using the new simulation scheme, we made the simulation realistic and thus useful for training purposes. Training a control policy has many significant advantages over manual controller design approach. Training can be easily automated and require much less effort than its alternative. This translates to a cheaper and faster development of a robot. Another advantage is performance: trained control policies have more diverse behaviors, and thus make the robot more capable of reacting to environmental changes.”
As can be seen in the above video, the robot is able to readjust its gait when it’s given a shove or boot. Usefully, in the event that it is knocked over completely, it’s also capable of getting back to its feet. This would make it more useful in a real world environment, and would potentially mean less human supervision as it carries out its tasks.
“This means that many of the tasks can be performed more reliably,” Hwangbo continued. “The weakness of the existing robots in terms in practical situations is their reliability. In case of a fall, a human operator has to intervene. This deterred industries from using legged systems. Our contribution makes legged systems more practical.”
ANYmal was recently put through its paces when it was used to carry out inspections on one of the world’s largest offshore power-distribution platforms in the North Sea.
- Leaps, bounds, and beyond: Robot agility is progressing at a feverish pace
- This basic human skill is the next major milestone for A.I.
- Disney Research’s robotic eyeballs have a freakishly humanlike stare
- To build a lifelike robotic hand, we first have to build a better robotic brain
- This groundbreaking new style of A.I. learns things in a totally different way