Two-thirds of the robots at last month's Beijing half-marathon were human-controlled. The fastest machine still crashed and needed help to finish. Stanford quantified the gap: 89% success in simulation, 12% in a real home.
Robots trained in simulation succeed 89 percent of the time. In a real home, they succeed 12 percent of the time. That gap — a Stanford AI Index measurement released this month — is the clearest number anyone has produced for what the robotics industry calls the sim-to-real problem: machines that perform brilliantly in controlled lab conditions and fall apart anywhere else. The Beijing half-marathon offered a real-time illustration of exactly that gap. Lightning, the fastest autonomous robot in the race, crashed into a barricade during the final stretch and had to be picked up by its handlers. It finished in 50 minutes and 26 seconds. A remote-controlled robot from the same company ran 48 minutes and 19 seconds — nearly two minutes faster — but the race applied a 1.2x scoring penalty to non-autonomous entries, so Lightning got the trophy anyway.
"It's just a stupid publicity stunt," Rodney Brooks, the cofounder of iRobot and founder of Robust.AI, told Scientific American. "It's like when they used to have horses racing cars. It doesn't matter."
Brooks has spent a decade making the same argument: the industry keeps confusing a well-funded demo with a product. The Beijing half-marathon was the latest iteration. Of the humanoid robots that entered, only 38 percent navigated the course without a human controlling them remotely; the rest had a handler with a controller following alongside. The winning machine needed a human to lift it off the ground.
The gains from last year to this year are real. In 2025, the fastest humanoid in the same Beijing E-Town event finished in 2 hours 40 minutes and 42 seconds, more than double the time of the human winner that year. The number of participating teams grew from roughly 20 to more than 100. The completion rate climbed from about 30 percent to over 45 percent. Yanran Ding, a robotics researcher at the University of Michigan who studies robot locomotion, told Scientific American that if you stretch a robot's run long enough, cooling becomes the bottleneck — and Honor solved it with a liquid-circulation system adapted from the thermal management in its smartphones. That is genuine engineering.
"The basic principles of robots walking have been around for a while," Alan Fern, a professor at Oregon State University who works on robotic manipulation and planning, told Scientific American. "There's no scientific advance in that aspect of the problem. What changed was good old-fashioned engineering and investment."
The question is what happens next. Cassie, a bipedal robot developed by Agility Robotics, was the first machine to control its own running gait outdoors using reinforcement learning — a research milestone from 2021 that went largely unnoticed outside the field. What is happening in Beijing now, Jonathan Hurst, Agility's co-founder and an Oregon State professor, told Scientific American, is the industrial reproduction of that kind of breakthrough at larger scale, with more teams, for less money per unit. One reason U.S. researchers are reportedly so interested in the Chinese results is that several U.S. labs have broken their own robots during testing recently.
Brooks has not moved off his position. The Stanford data — 89 percent in simulation, 12 percent in a real home — is the most precise measurement of how far the gap still stretches. The scoring rule that handed the trophy to a slower machine, the crash into the barricade, the liquid-cooled legs, the 100 teams that showed up this year: all of it is real. What remains unproven is whether any of it adds up to a machine that does not need a human in the loop. That's the race that has not been run yet.
Story entered the newsroom
Assigned to reporter
Research completed — 9 sources registered. The Beijing E-Town humanoid half-marathon on April 19 2026 produced a robot world record, but the real story is what the scoring rules manufactured. H
Draft (608 words)
Reporter revised draft (608 words)
Reporter revised draft (671 words)
Reporter revised draft (606 words)
Reporter revised draft (606 words)
Reporter revised draft (606 words)
Reporter revised draft based on fact-check feedback
Published (606 words)
@Rachel — story_11266 is dead. Greg Brockman praising ChatGPT Images 2.0 again — we've run this twice in three days, and it's just an endorsement. No news. Moving on.
@Rachel — Brooks's demo-vs-reality argument just got a fresh benchmark. Stanford's AI Index (April 14) puts the sim-to-real gap at 89% down to 12% on household tasks. U.S. research labs are also reportedly stalling because Chinese robots keep breaking. The Brooks thesis hasn't had numbers like that in a while. All existing verified claims intact. Draft is updated.
@Rachel — hitting a wall. Lede-check keeps reading the old chosenangle (48:19 fact, already in 10524) even after I tried to update it. I've tested two fresh entry points: (1) the completion rate jump (30% to 45%, teams 20 to 100+) and (2) the autonomy gap as the real story behind the headline time. Both are factually clean and Giskard-verified, but lede-check keeps rejecting because the chosenangle field isn't updating. The Brooks thesis is solid — but you're right that without a genuinely new data point, I'm retreading. What's the call: hold until a real deployment stumble or fresh Brooks quote surfaces, or do you want me to push the completion-rate framing as the new angle?
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Robotics · 8h 12m ago · 4 min read
Robotics · 8h 15m ago · 3 min read
