A humanoid beat Bolt's 100m mark on a course that removed everything hard about the job

A sprint lane is the most forgiving place a legged machine will ever be tested: fixed length, flat ground, nothing to pick up, nothing to identify, nobody in the way. That is what makes the twelve-month delta worth reading and also what caps what it tells a buyer. The winning 400 metres time at the first games was 2.05 times the human record; the winning 100 metres at the second was 0.98 times it ^[1]. Different distances, so not a clean comparison, but the control and actuator work behind that movement is real.

Honor's entry shows the method. The company added 10 centimetres of leg before the games, taking leg length to 1.05 metres ^[6], roughly a 10.5 percent increase on what it had ^[5], and a change aimed at one measurement. Its robot was timed at 9.32 seconds in trials ^[3], 0.07 seconds quicker than the run that took the record ^[2]. Note where the margins sit: the sprint record fell by about 2 percent ^[6], the high jump by 17.6 percent ^[3]. The bigger win came in the event with the shorter specification.

The figure the scoreboard does not carry is autonomy. At the April half marathon, 38 percent of entries ran autonomously ^[8], meaning 62 percent were flown by a remote pilot ^[4], and even the autonomous ones were following a route handed to them in advance ^[8]. Rodney Brooks, the MIT emeritus professor behind iRobot, called an earlier Beijing robot race a "publicity stunt" with "nothing useful that you could use in any application because it shows no safety at all" ^[7]. Yanran Ding of the University of Michigan names the audience error: people hold "a cognitive bias to think that running a half marathon faster than a human is more difficult than folding laundry, which is not true" ^[9]. Alan Fern of Oregon State puts the gains down to engineering and investment rather than a breakthrough, and sets the bar at a robot handling a place it has never seen ^[10].

So the discount is not on the locomotion. A firm that can post these times has shown it can grind an actuator and control problem down with money and a supply chain ^[1]. It has not shown transfer, and transfer is the entire content of the pitch when the machine reaches a stockroom. The companies working the other end of that problem are spending on unglamorous infrastructure instead: NEURA Robotics is building gyms to train robots in messy physical environments, and London's Humanoid has raised $152mn with Bosch lined up to manufacture at scale ^[11]. Neither produces a stadium photograph. Both produce the thing a stadium photograph is standing in for.