Last year's inaugural games produced a winning 100 metres of 21.50 seconds and a best standing high jump of 0.95 metres [6][7]. Measured against this year's marks, the sprint lost 12.11 seconds, a cut of about 56 percent in twelve months [1], and the standing jump came in roughly three times higher [2]. That is hardware iterating in public at a rate a procurement team should take seriously, and it is also why the wreckage matters: the machines setting those marks were the same class of machines coming apart on camera.
What the results table does not carry is which mode produced which number. The Verge's Robert Hart, who covered the five days, does not break the records down by autonomy level [18], so the 9.39 seconds is either a robot or a robot plus a pilot taking a scoring penalty, and a buyer reading the leaderboard cannot tell which. Records are a peak measurement by construction. A demo video is a best take by construction. Neither answers the question an operator has, which is what the thing does on the four hundredth repetition with nobody watching.
Two axes are worth more here than speed and height. First, is the task structured, with known geometry and a start signal? Second, does the robot close the loop, or does a person? Sprinting and jumping sit in the structured corner with the loop closed by the machine, and that is where the headline numbers came from. The unglamorous events sat in the other corner: soybeans lifted with tweezers, powder weighed, blocks stacked [14]. Those looked worse on television for the same reason they matter more, because the geometry changes every time and nobody rings a bell to start.
The teams building these machines are demonstrating peak capability under contest conditions, on a floor with crash mats and a crew. A warehouse needs something else entirely: mean time between interventions on one dull task, in a building with people in it. Those are different products at this stage, and the games are unusually honest about the distance between them, mostly because the failures happened in front of judges rather than in an edit bay.
So when a humanoid arrives in your building next quarter, skip the reel and ask for the maintenance log instead. How often does it fall on your task. How often does a person have to step in. And how many hours pass from the crate showing up to the first unit of useful work, which is the figure vendors are least eager to hand over. A vendor who answers with stadium footage has answered a different question, and the answer to that one is already public.
