Science2 distinct publishers3 min readPublished
The second World Humanoid Robot Games cut the machines' 100-metre time from 21.5 seconds to 8.64 in a year, while the dexterity events failed even with human operators in control gloves. That puts the bottleneck in the hands.
The Scientist · Science desk
Compiled by The ScientistSomething wrong?How this is made
Karen Liu's comparison is the one to keep. A sprint or a backflip is a trajectory: the machine knows every millisecond what torque to generate, and it does not have to think [8]. Track and field asks humanoids to do the thing their hardware already does well, turning stored energy into acceleration [8]. The Stanford computer scientist's point is that daily-life tasks are the hard ones, and that most people's intuition about laundry versus backflips runs backwards [7].
The nail-hammering event is where the diagnosis gets specific. Robots struggled to drive nails into a corkboard efficiently even when team members wearing control gloves were supplying the motion [10]. In that configuration a human provides the perception, the planning and the timing, and the task still fails. What is left is the hand. Liu's account of the mechanism fits: grippers cannot get enough contact with a tool to put strong force through it [9].
The athletics numbers, meanwhile, move fast. Same event, same company: 21.5 seconds last year, 8.64 this year [5][4], an improvement of 12.86 seconds, or 59.8 percent [1]. Average speed over the distance went from roughly 4.65 metres per second to roughly 11.57 [2]. That is a genuine rate of change and deserves the attention it is getting.
What the 8.64 does not tell you is what happens after the line. Scientific American reports that most of the 100-metre field crashed into the padded wall at the end, that one robot caught fire, and that a weightlifting entrant went into the judges' table [15]. The clock stops at 100 metres, so deceleration and graceful failure sit outside the measurement. Nature headlined the same five days as rapid progress with hurdles remaining [19]; Scientific American called them as ridiculous as they are impressive [20]. Both readings are available because the metric ends where the difficulty starts. Rodney Brooks, the MIT emeritus professor, gave the general version of this after a humanoid beat the human half-marathon record in April: watching a robot perform one thing fools us into crediting it with human-style general competence [16].
The organisers did build the part that would settle this. Scenario-based events, including shelving books and making beds, scored autonomous completion above teleoperation [3], which both concedes that the glove is on the field and gives you a way to price it. Booster Robotics, which brought nearly 300 machines, calls the games a real-world testing ground for embodied intelligence [11]. Boston Dynamics, which did not enter, states the same bar as humanoids ceasing to be toys [12].
One result cuts the other way. According to Scientific American, a humanoid played tennis against a human and anticipated unpredictable moves [17], and Nature describes a table-tennis showcase against a human Olympian [18]. A racket sport against a live opponent cannot be pre-scripted the way a jump can, which makes it the thinnest but most interesting evidence in Beijing. If next year's games publish per-task autonomy splits, that is the figure to read against an actual job description.
Ranked by verification strength, evidence, and original report placement.
The second annual World Humanoid Robot Games were held in Beijing and ran from 22 to 26 August.
The games drew more than 600 competing teams, mainly from Chinese universities, research institutions and robotics companies.
The games feature 51 events, including kickboxing, table tennis and soccer, with machines representing 16 countries taking part, including Germany, Japan and the United States.
A robot from Beijing company X-Humanoid won gold in the 100-metre sprint with a time of 8.64 seconds, almost one second faster than Usain Bolt's record-breaking race in 2009.
The same X-Humanoid robot won the 100-metre sprint at last year's games in 21.5 seconds.
Another X-Humanoid robot jumped 3.4 metres vertically from a standing position at this year's games, breaking a human record set in 1993, according to the state-run Chinese outlet the Global Times.
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · August 26, 2026
1 article · August 25, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
Humanoid robotics finally has a rubric that punishes the off-camera human1 distinct publisher
product
Unitree gave back $30bn and still sits near five times its pre-float valuation1 distinct publisher
product
Missile loading and suspect restraint move onto the robot product roadmap in Beijing1 distinct publisher
invest
Beijing fielded 2,000 humanoids the same month Washington banned importing them1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Two independent eyewitness accounts with named expert mechanism, but records rest partly on state media
Both publishers report the event first-hand with concrete, checkable figures (8.64 seconds, 21.5 seconds last year, 3.4 metres, 51 events, 16 countries, 600-plus teams, nearly 300 Booster robots), and the central bottleneck claim is attributed to a named domain expert at Stanford and corroborated by an observed task failure. Evidence is weakened by the standing-jump record and the hammering footage coming via state-run Global Times and CCTV+, by the absence of official result tables, and by single-sourced items such as the tennis match with no robot, score or trial count given.
Heavy competition participation, no evidence of real-world use
Adoption evidence in these sources is confined to the event itself: 600-plus teams, 51 events, 16 countries, and one vendor fielding nearly 300 machines, which is real and substantial as demonstration activity. Nothing in either source shows humanoids performing paid or sustained work outside the arena, and the tasks closest to real utility - hammering, and by implication shelving and bed-making - were not completed efficiently even with human operators. Boston Dynamics' absence and its 'toys' framing further indicate the field has not converted athletic results into deployment.
Record framing outruns the capability that matters
The circulating claim - humanoids beating Usain Bolt and a 1993 jump record - is true as reported yet measures the easiest class of motion, where trajectories are prescribed and no perception or planning is needed. The capability that would justify the broader 'robots entering everyday life' reading failed publicly in the same week, including under direct human teleoperation, and two credible voices (Brooks on single-task generalisation, Boston Dynamics on 'toys') say so explicitly. The gap is positive but not extreme, because both publishers themselves carry the caveats and one headlines the ridiculousness alongside the achievement rather than suppressing it.
Promotional and state-media channels sit close to the evidence channel
Two of the athletic records reach readers through state-linked Chinese outlets - the 3.4-metre jump via Global Times and the hammering footage via CCTV+ - inside an event hosted in Beijing and dominated by domestic vendors, giving the primary sources a direct interest in the progress narrative. Competing vendors also supply the framing: Booster Robotics, with nearly 300 machines entered, calls the games an important testing ground for embodied intelligence, while Boston Dynamics, which did not compete, is the voice cautioning against 'toys'. Both interested framings are disclosed in the sources rather than hidden, which is why this is not scored higher.
Solid on the core split, thin on anything commercial
Confidence is high for the central finding - fast scripted locomotion, failing dexterity, with the constraint in the grippers - because it is supported by two independent publishers, a named expert mechanism and a directly observed teleoperated failure. It is lower for magnitudes sourced to state media, for the unquantified tennis and table-tennis results, and for anything about commercialisation, cost or deployment, on which the cluster is silent.