CMU taught a robot dog to walk a balance beam

While discussing humanoid robots not long ago, someone told me their main issue with the form factor is that — from an evolutionary standpoint — we’re not built particularly well. That’s not to say that our bodies haven’t served us well, of course.

They’ve done the trick for a few hundred thousand years. It’s more that if you sat down with a talented product designer and asked them to whip up something from scratch, certain concerns would likely lead them in an entirely different direction.

Balance is on that list. Again, we’ve done just fine for ourselves, all things considered, but if balancing was high on your list of priorities, you might opt for something with four legs and a lower center of gravity.

This off-the-shelf dog robot is a perfectly fine place to start. The quadruped is plenty stable in its standard locomotion. As you’d likely expect, that changes quickly when you, say, stick it on top of a balance beam. That, however, is the sort of challenge you live for, if you’re a part of a lab like Carnegie Mellon University’s Robotics Institute.

“This experiment was huge,” says assistant professor Zachary Manchester. “I don’t think anyone has ever successfully done balance beam walking with a robot before.”

As to why this is such a big challenge… for starters, these robots aren’t designed to do it. Again, if you’re that omnipotent designer, you’d add in more flexibility and counter balance, for starters. The solution the team landed on is that big backpack you see in the above photo. That’s a reaction wheel actuator (RWA) — something used to help control the altitude of satellites.

“You basically have a big flywheel with a motor attached,” adds Manchester. “If you spin the heavy flywheel one way, it makes the satellite spin the other way. Now take that and put it on the body of a quadruped robot.”

CMU notes:

Manchester said it was easy to modify an existing control framework to account for the RWAs because the hardware doesn’t change the robot’s mass distribution, nor does it have the joint limitations of a tail or spine. Without needing to account for such constraints, the hardware can be modeled like a gyrostat (an idealized model of a spacecraft) and integrated into a standard model-predictive control algorithm.

Why, you may ask, would anyone spend time developing such a thing? Aside from the obvious satisfaction of watching a dog robot walk a balance beam, the most immediate answer is search and rescue. That’s long been a key application for these sorts of robots — sending machines where you wouldn’t normally send humans. It’s easy enough to see why balance is super important in just such a scenario.