It started with a track stand. Then a bunnyhop. Now? It’s a full front flip bunnyhop. The Ultra Mobility Vehicle (UMV) from the RAI Institute just landed one of the most absurd feats in freestyle bike history. And it doesn’t even have a rider.
Using Reinforcement Learning (RL), the AI-powered UMV has officially taught itself how to front flip bunnyhop.
“That’s exactly how I’d do it,” BMX and freestyle icon Ryan “R-Willy” Williams commented on the viral clip.
Reinforcement learning: trial, error and send
The RAI Institute still hasn’t responded to interview requests (likely too busy watching their own highlight reel), but their mission is crystal clear: build smarter, more agile robots.
The UMV uses RL: a kind of AI that learns by doing. No pre-programmed moves, no joystick in the background. Just a continuous loop of “try this,” “crash,” “adjust,” and “try again.” It’s basically a toddler with a carbon frame and infinite patience.
“We’re training all the policies in simulation,” one RAI engineer explains. “We try an action, evaluate the outcome, then reward or penalize that decision in future runs. The goal is 15 successful landings in a row. No flukes.”
They’ve now got a robot that can balance dynamically, react in real time and launch itself into tricks like a foam-pit rat at Woodward. Not bad for a machine riding what still appears to be a Specialized Hotwalk carbon balance bike.
More Crankworx than warehouse bot
Executive Director Marc Raibert says their broader goal is to build robots that “care for the disabled, free people from dangerous work and help people live better lives.”
That’s a noble goal. But this one looks like it’s training for Red Bull Joyride.
In their latest research video, RAI engineers describe the challenges: designing lightweight frames that survive repeated crashes, modeling real-world battery behavior to match simulation and teaching a robot to recognize curbs, tables or handrails.
“We want the robot to operate at human cadence,” they say. “Or superhuman.”
They’re not kidding.
The UMV is now hopping, flipping and navigating obstacles with balance and aggression most riders would envy. It even tracks its own center of mass in mid-air and adapts its body position for stable landings. Backwards manuals anyone? Yep. Robot’s got them on lock.
What’s next: bike flip or backflip?
This isn’t just a sideshow. The lab claims the technology powering UMV could have major implications for search and rescue, off-grid delivery or military recon. Basically anywhere balance, movement and autonomous adaptation matter.
But for now? It’s launching front flips and stacking views online.
You want to impress us, robo-ripper? Let’s see a nac-nac. A suicide no-hander. Or heck, just pull a clean 360.