Beijing Yizhuang Humanoid Robot Half-Marathon: How Did Embodied AI Score on Its Midterm?

On April 19, Beijing’s Yizhuang district hosted a special half-marathon.

The contestants weren’t human—they were humanoid robots. 26 brands, over 300 robots competing simultaneously, nearly 5x the scale of last year.

Honestly, this robot half-marathon was way more entertaining than I expected.

Let me give you the conclusion upfront: most robots didn’t run very well, but the significance of this competition wasn’t about how fast—it was about finishing at all.

I watched the livestream, and a few images really stuck with me.

Some robots fell right at the start, face-planting on the ground. Others ran out of battery halfway through and got carried off by staff. Some had bizarre walking postures, like they were doing the robot dance.

But there were some decent performances too. Unitree’s robot, for example—while not fast, stayed steady the whole way without falling. The Honor-Peking University joint team also did quite well.

What deserves the most attention in this competition is actually the roster. Over 60% of participating brands used Baidu Smart Cloud’s technical support. What does this tell us?

It shows embodied AI development is starting to form a layered ecosystem. The bottom layer is cloud platforms providing computing power and algorithm support; the top layer is robot manufacturers doing hardware innovation and scenario applications.

This model is actually quite similar to the smartphone industry—Qualcomm/MediaTek provide chips, Android provides the OS, phone makers do differentiation.

But embodied AI is far more complex than smartphones. Phones only need to process touch input; robots need to handle physical interaction with the real world. Every action—walking, turning, obstacle avoidance—behind it lies complex control algorithms.

From this competition, several trends emerge.

First, bipedal walking stability remains the biggest challenge. Many robots walk fine on flat ground, but as soon as they hit uneven surfaces or slopes, they show their true colors.

This indicates current control algorithms are still mainly based on idealized models, with insufficient adaptability to real-world environments.

Second, battery life is another bottleneck. A half-marathon is 21 kilometers, but many robots couldn’t even make it 10km before dying. This is a major problem in practical application scenarios—you can’t expect factory robots to stop and charge every hour.

Third, cost remains a constraint. While most participants were engineering prototypes, public information suggests a half-marathon-capable humanoid robot costs at least several hundred thousand RMB. At this price point, we’re still far from large-scale commercial deployment.

That said, this competition also showed me some positive signals.

First, both quantity and quality of participants have noticeably improved. Last year it might have been just standing up is enough; this year they’re already competing on who can run more steadily.

Second, industry chain collaboration is strengthening. Baidu Smart Cloud supporting so many teams this time shows underlying infrastructure is maturing.

Finally, and most surprisingly—audience reactions.

People used to think robots were science fiction. Now, seeing robots fall and get carried off, they instead feel pretty real. This shift in perception is important—people are starting to accept that robots aren’t perfect yet rather than expecting perfect robots.

My judgment: embodied AI is moving from lab demos toward real-world validation. There will be setbacks and failures along the way, but the direction is right.

As for when we’ll see truly practical commercial robots? My estimate is two to three years. Once costs drop below 100k, battery life reaches 8 hours, and stability hits 99%+, this market will truly explode.

This competition served as a midterm exam for the industry. The grade wasn’t excellent, but it passed.

How far do you think humanoid robots are from truly entering our daily lives?