Scientists Built a BMX Robot That Taught Itself How to Bunny Hop

I don’t know about you, but I’m over the whole robot thing. Sure, it’s neat when they climb stairs, play soccer, or shoot basketballs, but have none of these scientists and engineers seen Terminator? Don’t they know they’re actively building our inevitable doom? I wish these people would stop, but you know they never will. So if society must keep cranking out robots, they should at least be like this rad bunny-hopping BMX stunt machine.

Personnel at the Robotics and AI Institute (RAI) created a robot that can bunny hop higher than I ever could on a bike as a teenager. Ignoring my bruised pride, it’s an impressive achievement and, more importantly, really fun. It’s called an Ultra Mobility Vehicle (UMV) and isn’t designed to take anyone’s job away or be used in military operations (although, its tech will undoubtedly find its way to both eventually); it’s just a living BMX bike that can do tricks by itself.

To get the UMV to successfully hop, land, and balance on its own, RAI had to develop a new control unit that uses reinforcement learning, a style of machine learning where the software effectively rewards itself for successfully navigating its environment. It involves plenty of trial and error, but according to RAI, reinforcement learning allows both the team and machine to push past the boundaries of typical direct control and realize the ultimate potential of the hardware in question. For example, RAI incorporated reinforcement learning on a Boston Dynamics dog-like robot, and the model helped the robo-pup improve from its 1.6 meters-per-second top speed out of the box, to 5.2 meters-per-second.

It’s interesting to watch the UMV ride around in the video below. Sitting on its “seat” is that chunky control unit that’s attached to the bike via an articulating mechanical arm. Amazingly, that computer has no gyroscope or sensors for stabilization—it’s just a normal bike with a big heavy backpack, and it figured out the best way to speed up, slow down, and turn the front wheel to keep itself upright much in the same way humans do when we learn how to ride.

Observing the UMV deploy everything it’s learned is fascinating. To jump, the robot’s mechanical arm quickly stands the heavy control unit up as high as it can, which brings the whole bike’s momentum upward, and then that unit pulls the bike up into the air with the same arm. To dismount, it simply rides off the table but then, in midair, it once again extends the arm to reduce downward momentum, softening the landing. The speed and smoothness are as remarkable as the trick itself.

After its bunny hop, the robot even does a track stand, where it comes to a sudden stop and balances itself at a standstill. It’s just showing off at this point. Next, we get a little victory wheelie—cute.

I’m not sure what sort of practical applications this tech could be used for, but there are some helpful real-world use cases for this little guy. It could be a delivery vehicle in urban, traffic-dense areas, or be used on job sites to help carry heavy tools and materials. But the idea of machines learning how to extend the limits of their hardware is frightening if you’ve ever seen any sci-fi movie, ever. Let’s just hope it doesn’t become a machine gun-wielding death robot in 50 years.

Got tips? Send ’em to tips@thedrive.com

The post Scientists Built a BMX Robot That Taught Itself How to Bunny Hop appeared first on The Drive.