Beijing Humanoid Robot Half-Marathon: When AI Grows a Body, the Mechanical Evolution Just Began

April 19th, Beijing Yizhuang.

A group of robots stood at the half-marathon starting line. Human runners beside them. The scene felt surreal.

This was the world’s first humanoid robot half-marathon. Not a lab demo, not a few steps on flat ground—actually running 21 kilometers. With slopes, curves, wind, and dust.

The finishing robot was called ‘Tiangong,’ from Beijing Embodied Intelligent Robotics Innovation Center. I couldn’t find official timing data, but completing the race was already victory.

Honestly, my first reaction was: really?

Making robots run marathons is like having F1 cars compete on cargo capacity—mismatched scenarios. Isn’t robots’ greatest value in factory labor, home cleaning? Why make them run 21 kilometers?

But thinking deeper, I understood the race’s significance.

It’s not showing ‘how fast robots can run,’ but testing ‘robot survival capabilities in complex environments.’

Consider marathon track characteristics: uneven surfaces, uphills and downhills, wind changes, temperature fluctuations, sustained output for 2+ hours. Isn’t this a microcosm of the real world?

In labs, robots face controlled environments: flat floors, constant temperature and humidity, stable lighting. But the real world isn’t a lab.

This race forced engineers to solve practical problems:

• How to maintain balance on uneven ground?
• How to manage battery life for 21km without mid-race charging?
• How to handle emergencies like tripping or slipping?

Every problem is a necessary step for embodied AI from demo to product.

A VC friend in robotics said something accurate: ‘2026 is the ‘road-ready’ debut year for embodied AI.’

Previously, competition was about who could stand up, who could walk. Now it’s about who can adapt to real environments, who can work.

Besides Tiangong, this race featured products from Unitree, AgiBot, Fourier Intelligence, and other domestic players. Though most ‘withdrew,’ this同台竞技本身 marks industry maturation.

I noticed one detail: rules allowed battery swaps but not complete robot replacement. Very realistic—in real scenarios, robots can recharge/swap batteries but can’t ‘reboot.’

Another trend worth watching: the humanoid form factor itself.

Why humanoid shape? This debate has raged in the industry.

One camp argues it’s to adapt to human-built environments—door handles, stairs, chairs designed for humans. Robots need human form for seamless integration.

The other camp sees this as path dependency. If designing a robot-friendly world from scratch, wheeled bases plus robotic arms might be more efficient.

Both make sense. But from this marathon, ‘humanoid’ seems to be shifting from ‘showboating’ to ‘practical.’

Tiangong’s running gait is noticeably more natural than robots from a year ago. Cadence, stride length, arm swing—all closer to human. This means control algorithms are advancing.

However, I must pour some cold water.

Current embodied AI is still far from true ‘general-purpose robots.’

These marathon robots were likely optimized specifically for racing. A robot that can run 21km may not help you move a chair.

And cost issues remain unresolved. I estimate these competing robots have BOM costs above several hundred thousand RMB—still light-years from consumer markets.

Finally, and most crucially: intelligence.

These robots can run thanks to motion control algorithms, not ‘brains.’ They don’t autonomously decide where to run, how to avoid obstacles, when to accelerate or decelerate. These decisions remain human-preset.

True embodied intelligence requires ‘body plus brain’ integration. Bodies execute tasks; brains make autonomous decisions.

From this perspective, this marathon marks only a milestone in ‘physical evolution.’ ‘Brain evolution’ is still ongoing.

But regardless, this is a moment worth commemorating.

AI is stepping out from screens, beginning to possess physical form. This is a revolution longer than foundation models, and one that will more profoundly transform our world.

After all, intelligence that can change the physical world is real intelligence.