ANYmal is a truly remarkable robot, capable of standing and lifting things like a humanoid, or slinking around on all fours like a quadruped, with or without wheels. But what’s really surprised us now is the eerie grace it’s starting to move with.
ANYmal is the perfect name for this little fella, given its ability to shape-shift and switch modes depending on the task at hand. It looks crazily over-engineered next to a regular robot dog, but then it needs a ton of extra power and flexibility in its arms, legs, shoulders and hips when it’s time to stand up start doing useful work.
Late last year we saw it getting about on wheels in both bipedal and quadrupedal mode, and learning to chuck boxes into bins with a baggage handler’s degree of care for the contents within. It was also figuring out how to open doors. Check that video out:
Curiosity-Driven Learning of Joint Locomotion and Manipulation Tasks
Now, the ETH Zurich ANYmal team has been setting its locomotion AI to work on uneven terrains and obstacles. It’s back on stumpy feet again, instead of the wheels, and the team has been running it through a kind of parkour training focused on smoothy climbing over impediments and dynamically jumping or climbing down again.
“Before the project started, several of my researcher colleagues thought that legged robots had already reached the limits of their development potential,” said ETH doctoral student and parkour hobbyist Nikita Rudin, “but I had a different opinion. In fact, I was sure that a lot more could be done with the mechanics of legged robots.”
Through a trial-and-error process, ANYmal has learned to assess obstacles and decide on an approach, be it climbing over them, slithering under them, or whatever combination of movements have worked in the past.
And there’s something remarkable about watching the machine make these self-taught movements – a sense of an almost sinewy smoothness starting to emerge. A tightness and efficiency of motion that gives this hard, heavy, metallic beast a hint of cat-like grace.
ANYmal can do parkour and walk across rubble
The ANYmal team has combined these new machine-learning obstacle-handling capabilities with a more programmatic model-based control approach developed for an earlier project in which ANYmal was taught how to get around over piles of uneven rubble.
The result, says the team, is a solid ability to read and navigate something akin to a disaster area, plus a growing library of smooth movement patterns it can apply in a flexible manner when it decides they’re necessary. It’s now more sure-footed on slippery and unstable surfaces.
If and when it breaks out of the research phase and reaches the market, it’s likely to start out as a fairly pedestrian inspection robot for industrial and construction sites. But it’s one of the most interesting, flexible and capable transforming robots we’ve seen, and we’re fascinated to learn where it goes from there.
Source: ETH Zurich
