Today in Beijing, more than 300 humanoid robots are lining up to run 21 kilometers through the E-Town economic development zone. It is the second annual robot half-marathon, and this year the numbers are bigger in every direction: 112 teams (up from 15), five international entries, and for the first time, nearly 40% of the robots will navigate the course autonomously, relying entirely on their own sensors instead of a remote operator with a joystick.

On paper, this is a story about progress. Last year, the winning Tiangong Ultra model finished in 2 hours and 40 minutes. That is more than double the human winner, but it finished. This year, Tiangong Ultra will run fully autonomously, using large-scale simulation training to closely mimic a human gait. The race now includes paved slopes and parkland terrain instead of flat road.

But here is the part the Chinese government does not want you to focus on.

Social media footage of robots training in Beijing at night this month shows the full spectrum: some models successfully imitating human running at 14 km/h, others jerking awkwardly, some crashing into railings, and a few just falling over. The gap between the best and worst entries is enormous.

"The reason our applications are not taking off is that the robots' IQ is too low. The models are poor, their success rates are low," said Tang Wenbin, founder of embodied intelligence startup Yuanli Lingji, at a Beijing tech forum last month. "Honestly, the whole industry's level is still at a very elementary stage. Right now, a lot of what we see is dancing disguised as working."

That quote should be tattooed on every humanoid robotics investor deck in the world.

The Numbers Tell Two Stories

China dominates global humanoid robot installations, accounting for more than 80% of the 16,000 units installed worldwide in 2025, according to Counterpoint Research. The top US vendor, Tesla, accounted for just 5% of global humanoid installations. Market leaders AgiBot and Unitree each shipped more than 5,000 units last year. Unitree has pledged to expand production capacity to 75,000 humanoid robots annually.

Those are production numbers that should make every Western robotics company uncomfortable. But production and usefulness are different things.

According to Unitree's own IPO prospectus, its humanoid models are primarily used by research institutions, for dance performances, and as interactive guides in service establishments. Not factories. Not warehouses. Not the industrial floor where these machines need to prove themselves.

Georg Stieler, Asia managing director at the Stieler consultancy, put it diplomatically: "Humanoid robot makers need to find a balance between quality in products which are still under constant evolution and price pressure." Translation: they are shipping fast and figuring out the actual applications later.

The Data Play Nobody Is Talking About

Here is what actually matters about this race: it is not really about running. It is about data. Every robot on that course is generating hours of real-world locomotion data that feeds back into training models. UBTech, maker of the defending champion Tiangong Ultra, had fewer than 10 humanoids in factories in 2024. Last year it jumped to more than 1,000. This year, it plans to launch 10,000 full-size humanoid robots across commercial settings.

"When we talk about AI, it relies on how much data, especially high-quality data, we can collect," said UBTech Chief Business Officer Michael Tam. The race is a data collection exercise dressed up as a spectacle. And the 40% autonomous entry rate means these robots are generating something even more valuable: unscripted navigation decisions in the real world.

The Chinese government named embodied intelligence as one of its key industries for economic transformation. The half-marathon is the public face of that push. The factory floor is where it will actually be won or lost.

So watch the race today for the spectacle. But pay attention to the training footage. That is where the real story lives: some robots sprinting at human speeds, and others face-planting into guardrails. The gap between vision and reality in humanoid robotics has never been more visible.

Sources: Reuters, Indian Express, CGTN, Economist, Counterpoint Research, Digitimes.