Switzerland Develops Dog-like Robot ANYmal

ANYmal, a dog-like robot with four legs that can stand up after falling has been developed by scientists at ETH Zurich in Switzerland.

The ANYmal was first developed by scientists at the Robotic Systems Lab, at the Swiss Federal Institute of Technology in Zurich (ETH Zurich). Since established in 2016, it has turned commercial as part of the company and is still being developed.

The robot can go outside and solve problems in some dangerous places humans can’t go, including for search-and-rescue missions. ANYbotics also said the bot can conquer less-than-ideal conditions such as industrial sites, snowy landscapes, and woods.

The machine learning method called reinforcement learning, which involves letting the robot nearly learn and “think” in similar ways to animals, but with its own internal logic. In particular, in this method, robots would train themselves by trials and errors so it can figure out the best way to accomplish a task, walking for example.

However, due to the complexity of motions, it's not easy to apply reinforcement learning to a physical, legged robot. Lead author Jemin Hwangbo, a Robotic System Laboratory scientist at ETH Zurich, stated that so far, most scientists use computer simulations for robot learning. Yet, it's also difficult to use the data from these simulations in training real-life robots.

Developing control policies for sophisticated legged systems is extremely difficult, according to Hwangbo. He said via email that the number of situations robots face in reality is much larger than what control logic can cover.

But now, as in their new paper, Hwangbo and his team can create a neural network which makes it easier to translate simulation data to the robot. All these simulations are done almost one thousand times as fast as it is in the actual world. Plus, they can

The new training bot can keep itself standing still while researchers try their best to kick it. And now it can even overturn itself after a fall, which the research team says no other equally complex four-legged robots have ever done. In addition, ANYmal is now more energy efficient and faster, it can not only walk 25% faster than before but also better follow commands to move at a certain speed.

Then it took what it learned and applied it to the real world. It even learned how to run faster than before. Thanks to the neural network, ANYmal can also reach speeds of 1.5 meters per second or more than three miles in just a few hours.

Scientists used an AI pattern to teach ANYmal how to stand on either side or back in many cases, instead of just setting the robot to stand up in a particular type of environment. Hwango said, although what ANYmal can do now is a good thing, this robot has only learned from a flat surface. To overcome unstructured and rough terrain, the robot needs to have vision sensors and a suitable policy for processing information. They are working in this direction and hope to offer a more flexible solution soon, training robots to become more agile in many situations.