Humanoid Robots Run a Half-Marathon: The Ultimate Real-World Test

April 19, Beijing Yizhuang. Over 100 teams, 21 kilometers.

When the starting gun went off, I was watching the livestream. Honestly, my emotions were complicated—not as a spectator, but as someone in the industry watching peers run their first real gauntlet.

Glory’s “Qitian Dasheng” took the championship, finishing in 50 minutes and 26 seconds. Remember, that’s a half-marathon, not a 100-meter dash. For context, that’s middle-of-the-pack amateur runner level. Respectable.

But the more telling number is this: four in ten teams completed the course entirely on their own—no remote control, just algorithms and sensors navigating open roads.

That’s the real value of this competition.

Most embodied AI demos you’ve seen? 90% happen in controlled environments: fixed lighting, even floors, known obstacles. The real world doesn’t work that way. There are cracks in the pavement, tiles that buckle, delivery riders cutting across the path.

This half-marathon course threw all those real-world disruptions in, and let the robots figure it out. Pass this, and you’ve crossed a real threshold for “embodied intelligence.”

Of course, problems surfaced too.

One fact that gets overlooked: among those 100+ teams, only 40% finished autonomously. The other 60%? Either dropped out or needed human intervention. What does that tell us?

It means current embodied AI still has meaningful distance to cover on “sustained stable operation in open environments.”

But honestly, that’s a good thing. Knowing where the gap is matters way more than false optimism.

One detail I found telling: this competition attracted serious industrial capital. Not just Glory—Itashi, Yushu, other companies all sent teams. When an industry draws real money doing real engineering, that’s not pure hype.

Next checkpoint: same time next year. Let’s see how these “seed players” improve their times.