ROBNEWS
Robot Legs Learn a New Trick: Going Backward
The Robotics and AI Institute (RAI), founded by Marc Raibert, unveiled Roadrunner—a 15kg bipedal robot with no torso or arms, just wheeled legs capable of knee reversal steering.
The Robotics and AI Institute (RAI), founded by Marc Raibert, unveiled Roadrunner—a 15kg bipedal robot with no torso or arms, just wheeled legs capable of knee reversal steering.
Marc Raibert spent two decades teaching the world that robots could run, jump, and do backflips. Now his new shop has a video of a pair of robotic legs skating across a floor, climbing stairs, and balancing on one wheel — and it might be the strangest thing you'll see all week.
The Robotics and AI Institute (RAI), a Massachusetts-based research outfit that Raibert founded in 2022 with more than $400 million from Hyundai Motor Group, posted demo footage of a robot called Roadrunner on March 23. The 15-kilogram (33-pound) machine has no torso, no arms, and no particular reason to exist yet. It is, by any reasonable measure, a pair of robot legs bolted to two wheels. And it is genuinely mesmerizing to watch, according to New Atlas.
The trick is in the knees. Roadrunner's legs are fully symmetric — the same geometry top to bottom, left to right — which means they can point forward or backward with equal ease. To change direction, the robot simply bends its knees the other way. No differential drive, no spin-in-place maneuvers, no awkward pivot. It just reverses the bend and goes. RAI calls this knee-reversal steering, and per New Atlas, it means the robot can point its knees forward or backward to avoid obstacles or manage specific movements. It is a clean piece of mechanical ingenuity wrapped around a simple idea: if both orientations of the leg are equally valid, the robot has twice the room tomaneuver in tight spaces.
The legs also let Roadrunner switch between two wheel configurations on the fly. In side-by-side mode, the wheels track parallel, like a conventional rider. In inline mode, one wheel trails behind the other, narrow enough to thread through gaps a wider stance couldn't fit. The robot can lock its wheels and stomp up stairs, or leave them free and roll. It stands up from prone positions without external help. And it balances on a single wheel — one leg lifted, the whole machine steady, like a bike that forgot it had a kickstand.
What makes this mechanically interesting is not any single capability. Stair-climbing wheeled robots have been around for years. What RAI has built is a single system where all of these behaviors — rolling, stomping, balancing, stance-switching — share the same underlying control. According to RAI's own description of the work reported by Digital Trends, a single control policy was trained to handle both side-by-side and in-line driving modes, and multiple behaviors, including standing up from arbitrary positions and one-wheel balance, were deployed zero-shot on the hardware. That means the robot performed these tasks without per-task fine-tuning after the policy was trained — a standard benchmark in reinforcement learning for legged systems, but still uncommon in prototypes this fluid.
The zero-shot detail matters because it hints at what RAI is actually selling. Roadrunner is not a product. It is a research platform — a place to push the boundaries of what a single learned controller can manage across wildly different mobility regimes. RAI has said it plans to use the platform to explore agile, dynamic mobility for future applications in inspection and warehouse tasks, according to Yanko Design, though no commercial timeline has been announced. A more developed version, RAI suggests, could eventually serve as what it calls a viable alternative to leg-equipped humanoid robots in certain settings — faster than walking, more capable than pure wheels.
That framing is worth taking seriously, and worth skepticism in equal measure. Raibert's lab has a long track record of producing robots that look extraordinary in controlled demos and face a longer road to real-world utility than the footage suggests. Boston Dynamics spent years refining Atlas from a hydraulic powerhouse that could run and jump into a commercially available system, and even that system found its most reliable market in entertainment and inspection rather than the humanoid-labor displacement narrative that followed the humanoid wave of 2024 and 2025. The gap between what a robot does in a video and what it does in an unfamiliar warehouse, on an uneven sidewalk, in weather, remains the central problem in robotics that no demo has yet solved.
Hyundai's backing gives RAI the runway to chase that problem without immediate commercial pressure. The $400 million investment, confirmed in a 2022 press release from Hyundai Motor Group and Boston Dynamics AI Institute, is one of the larger standalone bets on pure robotics research in recent years. Whether it produces something that leaves the lab depends entirely on what happens next — and nobody, including RAI, is saying what that looks like yet.
For now, Roadrunner is a pair of legs, two wheels, and a video that makes you watch twice. That is not nothing. It is also not a product.