Enlarge image
Four-legged friend ANYmal wanders near the Etzel in Switzerland
Photo: Takahiro Miki / ETH Zurich
A mountain hike is often a small adventure - at least for most people living in the lowlands who only venture into the summit regions occasionally.
Robots are no different.
They usually spend most of their time in labs trying to master complex courses or do somersaults.
But now a research team led by Marco Hutter from the Swiss Federal Institute of Technology in Zurich (ETH) has sent their robot dog ANYmal into the wild.
On a hiking tour, the machine was supposed to climb the summit of the 1098 meter high Etzel at the southern end of Lake Zurich.
The fact that the tour was successful over steep passages and slippery ground is by no means a matter of course.
On the one hand, robots feel just like humans when they have to master unknown terrain.
We also feel our way around if we can't assess the ground exactly, so that we can move safely in difficult terrain.
We combine the impressions from the visual perception of our environment with the sensory impressions that our legs and arms feel.
Incidentally, we do this quite automatically and mostly unconsciously.
In addition, our experience teaches us which places are perhaps too difficult or dangerous and how we can better avoid them from the outset.
But marching robot legs only partially have this ability. "The reason is that the information about the immediate environment captured by laser sensors and cameras is often incomplete and ambiguous," says Takahiro Miki, a PhD student in Hutter's group and first author of a new study. That is, tall grass, puddles of rain, or snow can easily appear to the robot's cameras as insurmountable obstacles when the machine could actually traverse them.
There is another risk in the mountains that even experienced hikers know well: Difficult lighting conditions, dust or fog can cloud visibility.
While we can deal with such uncertainties, it may present the robot with a major task.
Because he now has less information about his environment and only an incomplete picture to make a decision.
The robot therefore had to learn to interpret and prioritize the images and impressions in order to then decide how and where to use the sense of touch on its legs on the tour. The researchers succeeded in doing this with a new control technology, they report in the specialist magazine "Science Robotics". Now he could learn to combine his visual perceptions with his so-called proprioception, the sense of touch. This enabled him to get through the terrain four minutes faster than the average hiker would expect. He effortlessly climbed the necessary 120 meters in a 31-minute hike – without any falls or missteps.
For the first time, a legged robot is now able to combine its external perception and its proprioceptive perception. To do this, ANYmal first had to go to a virtual training camp – a system based on a neural network, in which the four-legged friend learned in theory classes how to ideally overcome obstacles and rely on environmental data. "With this training, the robot is able to master the most difficult natural terrain without having seen it beforehand," quotes a statement from ETH Professor Hutter.
Incidentally, ANYmal is extremely prudent in the field.
When in doubt, the machine relies on its sense of touch and ignores vague sensor or camera data.
This makes it the ideal candidate for use after an earthquake, forest fire or a nuclear disaster, for example, wherever it is too dangerous for humans and where other robots cannot cope with the difficult terrain.
Or as a companion for lonely hikers.
joe