A humanoid robot beat the human half-marathon record at a Beijing race. But what did it actually prove?

Last Sunday, during the Beijing E-Town Half Marathon, a red humanoid robot Lightning type completed the course in 50 minutes and 26 seconds, faster than the human world record. Its long legs were modeled on those of elite runners, and its engines were cooled by a liquid circulation system adapted from smartphones made by its manufacturer, Honor, a Chinese phone company. A clip of the robot’s performance that ricocheted across the Internet was seen by many viewers as a milestone. It was, and it wasn’t. The Lightning robot also crashed into a barricade, fell and waited for its masters to put it back on its feet.

The Beijing race offered a striking insight into the state of humanoid robotics. Engineers have gotten much better at building machines that can travel long distances without overheating or breaking. Getting them to move in the real world with anything resembling human judgment is another matter.

This year, more than 100 teams registered, compared to around 20 the previous year, when the winning robot finished in two hours and 40 minutes. But only 38% of registrations for the 2026 event occurred independently, according to organizers; the others were controlled remotely. And all the robots followed a dedicated, rehearsed course, with support teams trailing behind.

On supporting science journalism

If you enjoy this article, please consider supporting our award-winning journalism by subscription. By purchasing a subscription, you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

“It’s just a stupid publicity stunt,” says Rodney Brooks, professor emeritus at the Massachusetts Institute of Technology, who co-founded iRobot, the company behind Roomba, and now runs the artificial intelligence-based robotics company Robust.AI. He has spent decades watching hype cycles around robotics come and go. “It’s like back in the day when they had horse racing cars,” Brooks says. “It doesn’t matter.”

His objection is not to the construction of Lightning or its era, but to the setting in which it was made. Lightning has not run a half marathon in human conditions. “There’s nothing useful that you can use in any application because it doesn’t show any security,” says Brooks. “There is no interaction with real people… and there is no possibility of interacting with the world because everything is pre-programmed,” he adds. “How many people supported [the robots] to have? How many vehicles followed them? People just do a marathon. They don’t know where the track is.

Brooks has been making a version of this argument for nearly a decade: Humans confuse performance with competence. “When you see how well an AI system or a robot performs on one thing,” he says, “it makes us believe that it has the same general skills as a human. And that’s a mistake people make.” If a human ran a half marathon, we would infer something about their balance, perception and resilience. The robot’s exploit does not generalize in this way. “If he ran into the crowd and it was safe, that would be amazing,” Brooks says. “But they are far from there.”

So what did the race actually measure? Alan Fern, a computer science professor at Oregon State University, has spent most of his career training bipedal robots to walk. With collaborator Jonathan Hurst, he helped build Cassie, a two-legged robot who ran an outdoor 5K in 2021 and later set a Guinness World Record in the 100 meters. The technique that allowed Cassie to function (training her in physics simulations) is likely also the origin of the abilities of robots like Lightning. “The basic principles of how robots walk have been around for a while,” says Fern. “There is no scientific progress on this aspect of the problem.”

What’s changed this year, Fern says, is “good old-fashioned engineering and investing.” Last year’s robots were slower and many broke; this year’s machines were fast and held up well. It’s not nothing, but it’s not a major breakthrough either.

When it comes to autonomy, Fern is gentler than Brooks, but not by much. Robots that operated without a human pilot always followed a route they already knew. “That meets the definition of autonomy for this particular task,” he says. “It’s a bit like we would talk about self-driving cars. The first self-driving cars could involve following a lane autonomously. The human doesn’t have to do anything and the robot will follow the lanes of the road.” Fern calls this specialized autonomy. What would have been scientifically interesting – and what wasn’t done with any robots at the Beijing event – ​​would have been to drop one off in “a whole new location” and ask it to navigate a crowded market, squeeze through restricted areas and avoid bumping into people.

Jonathan Hurst, who co-created Cassie and then co-founded Agility Robotics, sees the Beijing race as an inflection point in global interest in robotics. He believes that when Cassie ran her 5K in 2021, it was the first time a bipedal robot controlled its own running gait outdoors using reinforcement learning, a trial-and-error method in which an AI system is rewarded for successfully controlling its robot body. Few people noticed it. “Maybe there were just six people in the robotics community who said, ‘Holy shit, I didn’t realize reinforcement learning could actually control a robot,’” he says. Five years later, teams around the world are replicating this approach at a fraction of the cost.

This race to Beijing, according to Hurst, is not a single scientific leap. This is a field that reaches the scale where hard work can seriously begin. His own company has devoted much of the past two years to a specific problem: getting Digit, Agility Robotics’ iconic humanoid robot, to operate safely in a warehouse. This is the gap between a robot capable of executing a predefined route and a robot capable of moving safely among people. “It’s like looking at early cars and saying, ‘It doesn’t fly,'” Hurst says. “That’s a pretty high bar.”

The hardware improvements are significant, however. Yanran Ding, an assistant professor of robotics at the University of Michigan, views the race primarily as a hardware demonstration. “It’s really difficult to operate robots robustly for such a long period of time,” he says. The biggest feat was heat management. “The engine technology is there to work over short distances,” he says, “but if you stretch it, cooling becomes the bottleneck.” Honor’s design, Ding adds, addresses these limitations in the same way that a high-end desktop computer does. “Instead of using a fan, which uses air conduction to cool the chips,” he says, “you actually use a liquid circulation system: basically, put a [air conditioner] in the computer.

Ding also notes the design choices of the fastest robot bodies. “They have huge hip and knee motors. But they have very thin upper torsos: their arms are quite small but just big enough to provide inertial balance. And their shins and feet are also very light.” When running, he explains, you lose energy with each kick. “In order to minimize the energy loss that occurs when you hit the ground, you want to make your distal links as light as possible,” says Ding. In other words, Lightning was built like a greyhound, not a house cleaner.

“In terms of hardware, the limit is no longer the hardware: we can now really double down on the algorithms,” says Ding. “People, as humans…, have a cognitive bias that leads them to think that running a half marathon faster than a human is harder than folding laundry, which is not true.”

This reversal is really what the race demonstrated. A robot running fast seems difficult, because running fast is difficult for us. Folding a towel or moving safely through a crowded room seems trivial because we do these things without thinking. For robots, the hierarchy is often reversed. The muscles work. The brain, as Brooks pointed out, remains the problem.