Boston Dynamics Put Scooter Wheels on Its Humanoid Robot. It Nailed a Spinning Jump First Try.

Boston Dynamics took the feet off its humanoid robot, bolted on a pair of ordinary scooter wheels and let it loose in a mock warehouse. The robot, Atlas, slid sideways, balanced on one wheel with the other leg in the air, got itself up off the floor and then launched into a full 360-degree spinning jump. It landed it on the very first try.

It looks like a stunt for the internet, and it did rack up hundreds of thousands of views. But the company says the wheels were never the point. The point was finding out whether the robot’s motors and its SI brain can be trusted when everything is pushed to the edge.

A bicycle-wheel swap

The hardware change was almost comically simple. According to Boston Dynamics’ own write-up, a single custom adapter mounts an off-the-shelf scooter rim onto the robot’s existing ankle motor. That adapter is the only new part in the whole build. Because Atlas was designed from the start to be built in high volume and to be easy to service, the swap was “about as involved as swapping a wheel on a bicycle,” the company wrote.

The software change was just as small. Each wheel is treated as an ankle that can spin forever. Engineers took the standard model of the robot, swapped the feet for wheels and retrained.

Why wheels at all? Because rolling around under real weight is brutal on motors. Every launch, turn and stop forces them through their full range of torque and speed while carrying the robot’s load the whole time. That makes it a perfect stress test, the kind of punishment a robot might face on a factory floor that a lab can’t easily recreate.

Practice in a computer, perform in the real world

The SI behind the moves was trained with a method called reinforcement learning. In plain terms, the robot practices millions of times inside a computer simulation, gets rewarded when it does well, and slowly figures out what works. Only then does the finished “policy” get loaded onto the real machine.

The catch is that the simulation has to match reality. If the virtual motors are too perfect, the robot learns tricks its real body can’t pull off. Engineers call that the sim-to-real gap, and closing it is one of the hardest problems in robotics.

Andy Wu, a controls engineer on the Atlas team, said a lot of the work was “unglamorous but important”: checking settings across every piece of software, running tests on the real motors and tuning the simulated ones until they behaved the same way. Once that was done, he said, the moves trained in simulation transferred to the hardware “basically perfectly.”

That is the real headline for anyone who cares where robots are going. Teaching a robot in a computer is cheap and fast. Teaching it in the real world is slow, expensive and breaks things. The closer a company can get the two to match, the faster it can teach robots new jobs.

The spin was the easy part

Here is the twist. Wu said the 360-degree jump was one of the lower-risk moves, even though it looks the most dramatic. The robot was following a recording of a real person doing the jump, captured with motion-capture gear, and the training pipeline simply learned to track it.

The move that surprised him was the side dash. A person doing that motion would slide their feet. Nobody told the robot how to handle that on wheels. During training it figured out on its own that it could steer the wheels and roll through the slide instead.

“I was pleasantly surprised that training did, in fact, figure out how to use the wheels to roll during the side dash,” Wu said in the company’s video. He also admitted the project was a blast. Driving the robot around, he said, is “like playing a video game,” and he hasn’t touched his gaming rig since he started.

Even before the jump, the wheeled Atlas behaved differently from the walking version. The company says that right before takeoff, its lower body quickly rolled into position to set up the push, a small adjustment it worked out for itself.

Don’t expect a rolling Atlas in stores

Boston Dynamics was clear that a wheeled Atlas is not a product. It is a research testbed, and the company has no plans to sell it, according to The Korea Times. The company, which is owned by Hyundai Motor Group, says years of data from its Spot robot dog and its Stretch box-unloading robot taught it that machines in the real world run into problems a lab can’t fully recreate.

It also isn’t the first time the company has played with wheels. Nearly a decade ago it introduced Handle, a robot that balanced on two wheels, as The Verge pointed out.

Wu said his next project tackles a question that matters far more to anyone who might someday work next to one of these machines: what the robot should do when one of its motors fails. He wants to see whether the same training approach can teach the robot to react on its own and keep going, rather than simply crashing to the floor.

Measured against that, the scooter wheels are a fun sideshow. What Boston Dynamics actually showed is that its virtual robot and its real robot are getting close enough to trust, and that the SI can still find tricks nobody programmed. For a company that wants humanoids working shifts in factories, that matters more than any spin.