The second World Humanoid Robot Games closes today in Beijing, and it will be remembered for a single number: 9.39. That is the time a humanoid robot posted over 100 meters at the National Speed Skating Oval, two tenths of a second under Usain Bolt’s 2009 world record of 9.58. AP reported that robots also broke the human standing high jump record on the same opening day, before hundreds of humanoids marched in synchronized formation across the field.

The other images from these games are less flattering. An Honor-built robot disintegrating mid-race in a shower of components next to two Tian Gong Ultra runners. Robots crossing the finish line and immediately crashing into barriers because they cannot decelerate. Fallen machines snapping at the waist in bursts of sparks. At least one unit catching fire on the track.

Both sets of images describe the same industry, and both are accurate. The mistake would be treating either one as the whole picture. The games, taken seriously as a dataset rather than as a highlight reel, tell us three things: whole-body control has made a genuine one-year leap, task-specific performance still dominates general competence, and the commercial and political environment around humanoids is now changing faster than the hardware.

The scale of the Thing

First, the raw numbers, because they matter more than the viral clips. The five-day event (August 22-26) hosted 2,056 humanoid robots across more than 600 teams, 51 events and 1,301 matches. This is roughly triple the robot count of the inaugural 2025 edition, and it ran concurrently with the World Robot Conference elsewhere in Beijing, where companies exhibited around 3,000 products.

Two thousand humanoids in one venue is a statement about supply chains as much as about AI. China manufactures the majority of the world’s humanoid robots, and the games function as a national capability demonstration: actuators, harmonic drives, lithium packs and policy networks, all produced at scale and assembled into bodies that can survive, mostly, a week of competition. No other country could currently field this field, which is precisely the point of hosting it.

New events this year included table tennis and kickboxing, alongside returning soccer, dance and combat competitions. A Beijing company called Galbot staged what it described as the world’s first human-robot tennis match, with a robot that tracked balls, shuffled sideways, returned forehands and, when it lost balance chasing a drop shot, toppled backwards, rolled over, stood up and resumed play. That recovery sequence, unglamorous as it sounds, may be the most commercially relevant clip the games produced.

What 9.39 Seconds Does and Does Not Measure

The sprint record is real engineering. Dipam Patel, a PhD candidate at Purdue and a research assistant at the US Army DevCom Army Research Lab, put it to Ars Technica plainly: sprinting at that level requires coordinating every joint in the body simultaneously, and “there has definitely been a huge push in engineering from where they were last time to this time.” Going from trailing human records in 2025 to beating Bolt’s mark in 2026 is not camera trickery. It is a measurable annual delta in actuator power density, gait control and sim-to-real transfer.

It is also a narrow measurement. The running robots were purpose-built specialists that sprinted in a straight line and, in several cases, could not complete a controlled stop after the finish. The fastest machine in Beijing has no opinion about your warehouse. Companies fielded different robots for different events, each engineered to win one thing, and the scoring rewarded exactly that single-task peak.

The most instructive race may have been the small-group 400 meters, won by an X-Humanoid machine that ran with its arms raised beside its face in what viewers described as a “shy person hiding behind their hands” posture. The gait was not designed. It emerged from reinforcement learning in simulation, which discovered that swinging arms through hip rotation beat a human-style arm carriage for this particular morphology. That is sim-to-real working exactly as advertised, and it is also a reminder that the resulting behaviors are optimized artifacts, not general intelligence. The robot found the fastest way around a track. It did not find a way to slow down.

The Boring Events Are the Real Benchmark

The most forward-looking change at this year’s games was not on the track. Organizers expanded scenario-based competitions into hotels, homes, logistics and industrial production, and these events were scored on time and task completion rather than spectacle.

The menu, per Ars Technica and Global Times reporting: washing and hanging laundry, folding clothes, tidying a living room, and handling an unscheduled interruption in which a delivery arrived and the robot had to switch tasks to accept the package. At the Beijing Continental Grand Hotel, timed challenges included moving wheeled luggage to assigned rooms, stripping and remaking beds, and restocking supplies. A library event tested visual recognition and reasoning through book returns and correct reshelving.

The headline result from the scenario track: robots still lag typical human working speeds on household chores, and the hardest event was sobering. In the outdoor firefighting and rescue challenge, robots had 30 minutes to identify hazardous substances, locate and shut off three different valve types, spot a fire, and extinguish it. Three of twelve teams finished.

There is also an asterisk that every buyer should internalize: many scenario competitors were teleoperated by human operators. The rules allowed it, applying a scoring penalty rather than disqualification. This is a fair way to run a games, and a terrible way to read a demo. A teleoperated humanoid folding laundry is a puppet with good hands. The penalty system at least makes the games more honest than most corporate demo videos, where the teleoperator is never mentioned at all.

The Tether Problem

Beneath the autonomy question sits an infrastructural one. The South China Morning Post reported that even the more autonomous competitors relied heavily on cloud compute, with 5G modules streaming sensor data to an off-board “embodied-intelligence system” that returned instructions. The intelligence, in other words, was partially hanging from a cellular network.

This is not cheating; it is a rational engineering trade under current onboard power budgets. Vision-language-action models large enough to generalize are heavy, and running them on a robot’s internal compute drains batteries that already limit useful work to a few hours. But it defines the deployment envelope. A humanoid that needs a low-latency link to a datacenter is a humanoid that works in stadiums, convention centers and owned facilities with dense 5G, and works less well in the basements, rural sites and structurally noisy environments where much of the actual labor is.

The commercial race of the next two years is less “can a robot run 100 meters” and more “can a robot run its policy locally, all shift, on battery.” That race has no medal ceremony.

The Games Next Door to a Ban

While Beijing staged 2,056 humanoids, Washington was busy making sure most of them never enter the United States. In July, the FCC banned imports of new models of foreign-made robots on national security grounds, a rule aimed squarely at China that also swept in foreign robot vacuums. The Pentagon has added Unitree, the games’ most familiar brand, to its list of companies with alleged Chinese military ties. Legislation including the American Security Robotics Act and the GUARD Act is moving through Congress to harden the policy into law.

The market response is already visible. RoboStore, the largest North American distributor of Unitree hardware, which says it sold more than 1,500 robots to customers including OpenAI, Nvidia, Cisco and over 150 universities, has shut its distribution business. In August it launched Robo Inc. and is standing up a 66,000-square-foot manufacturing facility on Long Island, targeting commercial production by the first fiscal quarter of 2027. Its founder’s candor is notable: a robotic arm from Standard Bots or Universal Robots “is better suited for most businesses, for what they want to achieve, than a humanoid robot.”

So the world’s humanoid market is now formally bifurcated. On one side: Chinese platforms, iterating publicly at competition speed, getting cheaper and more capable, and selling everywhere except a closing American market. On the other: a US ecosystem whose newest manufacturing entrant is a former distributor, whose first products include a wheeled humanoid torso for healthcare rather than a general-purpose biped, and whose near-term business case is regulatory as much as technical. Investors poured more than $6 billion into humanoid companies in 2025 alone; the policy environment now decides which half of that capital can reach which hardware.

What a Real Benchmark Would Look Like

The games’ organizers deserve credit for the scenario events, because sport alone cannot price a robot. But the scoring still rewards peak performance over reliability, and that is backwards relative to what industrial customers buy. Patel’s proposal to Ars is the right one, cheap to implement and devastating if applied: make the robots run the same race ten times. A machine that posts 9.39 once and catches fire twice is a prototype. A machine that posts 10.5 ten times and stops cleanly every time is a product.

Extend that logic to the scenario track and you get the industry’s real to-do list. Repeat the laundry event across 50 unseen households. Run the valve task without the 5G tether. Publish mean time between failures and mean interventions per shift. Score teleoperation out of the result entirely, not with a penalty. None of these numbers would trend as well as a sprint record. All of them would tell a facilities manager more than the current highlight reels do.

The uncomfortable synthesis is that the two images from Beijing, the record and the fire, are not in tension. They are the same finding reported twice: purpose-built humanoids have crossed into superhuman territory on narrow tasks, while general-purpose deployment remains bounded by stopping ability, onboard compute, battery life and the cost of teaching each new task. The 9.39 is a genuine milestone in whole-body control. The smoke at the finish line is an equally genuine audit of how narrow that milestone is.

Next year’s games will add events, and records will fall again. The number worth watching will not be the 100m. It will be how many robots finish the boring events, untethered, unteleoperated, ten times in a row.

Sources