Atlas on wheels: Boston Dynamics' tests bipedal wheeled robots

Atlas, dancing on wheels

Boston Dynamics' Atlas robot did something unexpected this week. Engineers removed the feet, bolted on electric scooter wheels, and watched their humanoid robot land a 360-degree spinning jump on the very first real-world attempt.

It also taught itself to drift.

Unfortunately, this wasn't a product launch, at least not yet. It was engineering research, done in an creative way. I'm betting the wheels show up in a future edition of Atlas, though, similar to a fellow wheeled biped from Hexagon Robotics, Aeon.

The swap was simpler than expected

Engineers attached standard electric scooter rims to Atlas's existing ankle actuators using a single custom adapter. That's the entire hardware change. On the software side, the team swapped out the simulated feet in the robot's existing model for wheels capable of continuous rotation: no major rebuild, no separate platform.

Controls engineer Xiaoyu Wu described the hardware side as simply "just swapped a bicycle wheel."

What that tells you about Atlas's design is that the ankle actuators are powerful enough to serve as wheel motors under real operating loads. That's a margin built into the hardware, likely, that the team was able to exploit without modification.

What the robot did

The wheeled Atlas demonstrated rapid sideways movement, directional changes, single-wheel balancing while lifting the opposite leg, and fluid floor-to-acceleration sequences.

Interestingly, the sliding and drifting behaviors emerged without anyone programming them. The reinforcement learning policy discovered those movements on its own, finding ways to exploit available traction during rapid direction changes. That's a detail that surprised the engineers: trained policies developing strategies that weren't in the original specification.

The 360-degree jump required precise launch timing, of course. The controller tracked a trajectory recorded from an actual human jumper, with Atlas rolling its lower body into position before takeoff. It stuck the landing on the first try, which is impressive.

"I was pleasantly surprised that training did in fact figure out how to use the wheels to roll during the side dash," says controls engineer Andy Wu.

A research testbed

"A wheeled Atlas isn't a product," said Wu. "It's a research testbed."

The goal was to push Atlas's actuators toward their limits in a way that standard walking can't achieve. Bipedal locomotion places certain demands on motors: intermittent, high-force ground contact, with built-in recovery time between steps. Wheels change that profile entirely, requiring motors to deliver high torque at sustained high rotational speeds, continuously, while also supporting the robot's full weight through acceleration, turning, and braking.

That's a stress profile that normal walking doesn't have.

The technical approach layered classical model-based motor control for stability with whole-body reinforcement learning trained in simulation using Google DeepMind's MuJoCo engine. The control policy converted proprioceptive feedback and velocity commands into real-time pose and force adjustments across the entire body.

Closing the sim-to-real gap

Getting RL-trained policies to transfer cleanly from simulation to physical hardware is one of robotics' harder problems, for obvious reasons. Simulations are approximations, motors behave slightly differently from their models ... floor friction is never quite what the physics engine predicts.

Boston Dynamics addressed this by running physical experiments to measure how the actual actuators behave, then adjusting simulation parameters to match the measured data before training began. The payoff was policies that transferred to the real robot without the usual fragility.

The proof-in-the-pudding was the first-try success on the 360-degree jump.

The team is publishing these findings and plans to apply the results back to Atlas's standard legged configuration. The goal: improving reliability on 2 feet by understanding limits on 2 wheels first.

What comes next

The team wants to understand what happens when an actuator fails mid-motion. Can the whole-body control policy compensate? Does the robot stay upright if a wheel motor stutters under sustained load?

Those are the kinds of edge cases that determine whether a robot is reliable enough for real deployment, not just impressive in a lab video.

Meanwhile, Hyundai, which owns Boston Dynamics' parent, is expanding Atlas testing at its manufacturing facilities. Wheels could be very helpful indeed in massive vehicle production factories.

Sources

• Boston Dynamics blog post
• X
• MLQ.ai — Boston Dynamics' Atlas scooter wheels
• Korea Times — Atlas 360-degree spinning jump
• Korea JoongAng Daily — Atlas is on a roll