Beijing's robot games: the tweezers matter more than the 8.86-second 100 metres

The 100 metres is the event the current toolkit was built to win. Compact, high-power motors that hold a machine balanced at speed only became available in the past few years ^[14], and the gaits themselves come from policies trained in simulation, which is why several of the runners move in ways no human would ^[13]. Simulation is generous to legs. Ground contact is short, it repeats thousands of times per training run, and a bad friction estimate costs the policy very little. Hands are where the generosity stops: friction and slippage are precisely the quantities that are hard to predict ^[11], and a grasp that slips does not get averaged away by the next stride.

That is the mechanism behind the line Stefanie Tellex of Brown University gives Wired. Backflips look harder than plugging in a cable and are in fact easier, which she calls an example of Moravec's paradox ^[10]. Manipulating objects inside a more complex environment is far more difficult than sprinting or leaping ^[9], and advanced manipulation remains the field's standing unsolved problem, with some US startups going at it through more capable AI rather than better actuators ^[19].

Which is why the boring events are the informative ones. Putting a cable into a socket, and moving a piece of rubbish into a bin in a mock bedroom ^[8], are the first items on the programme that resemble work someone would pay for. They also sit outside the area where the host country's advantage is best documented. Tellex's own summary is that China has had amazing success building robot hardware ^[15], and Unitree, out of Hangzhou and recently listed, made its name on exactly those hardware gains ^[16]. A schedule that rewards contact-rich fiddling moves the contest onto ground where motor power is not the binding constraint. The same games had robots blowing out motors because they could not stop cleanly ^[18], which is a fair marker of where the mechanical envelope currently sits.

The caveat travels with the result. Some images and video appear to show humans remotely operating some of the robots, which suggests the machines were not handling those tasks autonomously ^[12]. That does not void the dexterity events, but it changes what they measure: a teleoperated cable insertion scores the hand, the wrist and the latency of the link, not the policy driving them. Wired's reading is that hands deft enough to grasp tiny objects with tools, even under human control, is still a real result ^[26], and that is defensible as a hardware statement and worthless as an autonomy one.

Scale is what turns a schedule into an agenda. More than 600 teams brought over 2,000 robots ^[2], roughly 3.3 machines per team ^[20], drawn from a domestic base of more than 100 humanoid manufacturers plus entrants from 15 other countries ^[3]. When a field that size optimises against one event list, the organisers are effectively writing a shared specification for next year's R&D. The thing the source does not tell us is how the manipulation events were scored, or which entries were remote-driven. Until someone publishes that, the 8.86-second sprint ^[4] is the only number from Beijing you can actually take at face value, and it is the one that says least about deployment.