The World Humanoid Robot Games began a five-day run in Beijing on Saturday, with more than 100 teams competing and 47 of them completing a robot half-marathon. Xiaomi, Unitree, Honor and Boston Dynamics are among the entrants. A Unitree H1 crossed the line with no engineer walking alongside it, and Honor's Flash finished in 50 minutes and 26 seconds.
Sixteen months ago, at the April 2025 event, many machines collapsed, hit barriers or did not finish. That is a real improvement over a short period, and it is worth being precise about what improved.
What a half-marathon measures
Not intelligence. A robot covering 21 kilometres on its own legs is a test of battery energy density, of thermal management when motors run hot for an hour, of joint actuators surviving several hundred thousand load cycles, and of balance control that degrades gracefully as components heat and drift.
Those are materials and mechanical engineering problems, and they are exactly the problems Chinese manufacturers are best positioned to solve, because they are the problems that yield to iteration, supply chain depth and cost reduction rather than to research breakthroughs.
It is also the right thing to test first. A machine that cannot stay upright for an hour cannot do a shift in a warehouse, whatever its software can reason about.
The tasks that matter more
The Games also judge robots on tightening screws, opening bottles and picking up beads with tweezers.
Those sound trivial next to a half-marathon and they are considerably harder. Walking is a closed-loop control problem with a well-understood objective: stay upright, move forward. Manipulation requires estimating an object's position, shape, weight and fragility from imperfect sensing, choosing a grasp, applying force accurate to within a fraction of a newton, and recovering when the object slips.
Humans do this without attention. Robots fail at it constantly, and the failure modes are expensive: a dropped part, a stripped thread, a crushed component.
This is why locomotion demonstrations have been impressive for a decade while commercial deployment has stayed narrow. The barrier to a humanoid doing useful work was never the walking.
The price gap
The number that should hold an investor's attention is the comparison between Xiaomi's Cyberdog at about $1,500 and Boston Dynamics' Spot at more than $75,000.
Both are quadrupeds, so this is a like-for-like comparison, and a fiftyfold difference is not explained by margin alone. It reflects two different theories of the market. Spot is sold into hazardous inspection work at oil and gas facilities, substations and disaster sites, where the machine's value is measured against sending a person somewhere dangerous, and where a five-figure price is trivial next to the alternative. Cyberdog is priced where the size of the addressable market is set by the price itself.
Whether the cheap machine eventually eats the expensive one depends on whether reliability at the low end reaches the level industrial buyers require. That is the same question that decided solar panels, lithium cells and drones, and in each of those the answer eventually favoured the cheaper manufacturer.
What the state has put behind it
Robotics has been formal Chinese industrial policy for over a decade. Xi Jinping called it the "crown jewel" of advanced manufacturing in 2014, Beijing hosted the first World Robot Conference in 2015, and a 2021 state plan attached subsidies, research tax breaks and financing to the sector.
An event of this kind is partly a demonstration to that policy's authors that the money is producing results, and partly a signal to foreign buyers and investors. Both audiences are intended.
The proportion to keep
Unit volumes in humanoid robotics remain small, and the projections being quoted for the 2030s rest on the manipulation problem being solved on a schedule nobody can currently evidence. Boursel has reported separately on Ant Group's move into humanoids and on Morgan Stanley's work on the Chinese sector, and the same caution applies to both.
What changed this weekend is that a category which looked like a research exercise in April 2025 now looks like an engineering programme with a falling cost curve. What has not changed is that nobody has shown a humanoid robot doing a day's paid work at a price a customer will pay, repeatedly, without a technician nearby.
The screws and the tweezers are where to look for that. Not the finishing times.
