The week produced two humanoid prizes, and they do not designate the same machine. The first is awarded on an athletics track in Beijing : it rewards legs, which beat two human running records in two days. The second is signed at the bottom of a purchase order : it goes to the quadruped ten times more often than to the biped, and it pays the biped more than sixteen times the price of the quadruped. Same week, same manufacturers, opposite verdicts.
The rules require autonomous decision-making where there is nothing to decide
The World Humanoid Robot Games assign a coefficient of 1.0 to tasks executed in full autonomous decision-making, versus 0.5 when they are teleoperated (Xinhua). This slider separates the proven from the assisted, and it does not apply everywhere : according to Forbes, the only media outlet to detail this point, flat races impose total autonomous decision-making, while hurdles, jumps, weightlifting and martial arts still tolerate teleoperation.
The strictest certification therefore applies to the event where there is nothing to decide : running straight down an empty lane. This is where the records fell. A humanoid covered the 100 metres in 9.39 seconds, under Usain Bolt’s 9.58 seconds in Berlin in 2009 (Euronews) the Tien Kung Ultra from the Beijing Innovation Center of Humanoid Robotics, timed at 9.32 seconds in the heats (Bangkok Post), took the 400-metre final in 38.15 seconds, under Wayde van Niekerk’s 43.03 seconds (Global Times). The locomotive leap is real : the winning time for the 100 metres was 21.50 seconds last year, meaning a speed more than doubled in twelve months.
The 2.88-metre high jump — against 0.95 metres last year and 2.45 metres for Javier Sotomayor in 1993, according to ESPN, the only source to report the mark — belongs, however, to a family of events where the rules do not impose autonomous decision-making. The figure is spectacular nothing obliges the machine to have achieved it without a human in the loop.
The drop is sharp as soon as it comes to chaining
The programme went from 26 to 51 events, including 21 professional or domestic scenarios (Xinhua) : the organisers wanted to measure the task, not the gesture. On the domestic scenario — tidying the living room, folding and putting away laundry, loading the washing machine, hanging out the laundry, receiving a parcel — accompanied by a 30-minute limit, no robot completed the entirety of the tasks during the period observed by the ZDF. On the fire-fighting and rescue event — identifying two hazardous substances, closing three randomly selected valves, extinguishing a simulated fire in 30 minutes —, only three teams out of the 12 competing that day went to the end, out of 23 registered (Interesting Engineering).
The same gap can be read within athletics : the long jump gold went to 7.97 metres, but out of 21 teams competing, only six recorded a valid attempt — a figure reported by the sole Bastille Post. And several sprinters did not know how to brake or correct their trajectory after the finish line : one of them crashed into the cushioning mat, fell backwards and caught fire (Bangkok Post). Machines that run the 100 metres under ten seconds and ignore what comes after the line do not have a speed problem, they have a control loop problem. The competitors were 96 % Chinese (Forbes) : this is what the densest industry knows how to do today.
The State Grid envelope says the opposite of the podium
On 19 August, 36Kr detailed the centralised purchase of State Grid Corporation of China : 6.8 billion yuan committed in 2026 for 8 500 embodied AI equipment. The envelope is split into 5.8 billion yuan for equipment, 800 million for research and development and 200 million for training, the purchase being phased in three waves — pilot in the first quarter, large-scale purchase in the third, complement in the fourth. The hardware fits in three lines : 5 000 inspection quadrupeds for 1.5 billion, 500 live-line work humanoids for 2.5 billion, 3 000 two-armed inspection robots for 1.8 billion.
Added together, these three lines make exactly the ‘ equipment ’ line. The arithmetic closes, and the average prices per unit it implies — calculation by the editorial team, State Grid publishing no tariffs — give 300 000 yuan the quadruped, 600 000 yuan the two-armed robot, 5 million yuan the humanoid. A humanoid is thus worth 16.7 quadrupeds : 5.9 % of the machines ordered, 43 % of the hardware bill. Usage reservation, a centralised purchase line can cover integration, software and deployment in addition to hardware. The ten-to-one ratio, however, does not depend on any calculation : it is written in the order.
The distribution of missions says why. Humanoids are reserved for live-line work on distribution networks and very high-voltage sites quadrupeds cover substations, transmission lines and networks in mountainous areas two-armed robots, the manoeuvring of substation equipment and fault handling. The biped is not bought to move, but for the rare gestures where a human torso and two arms are the only possible geometry — and billed accordingly. Deep Robotics, Unitree, AgiBot, UBTech and Fourier Intelligence are designated as the main expected beneficiaries (Humanoids Daily).
What enters the factory is almost never a biped
Bosch will launch the mass production of humanoids in August 2027 at its site in Bühl, in the Black Forest, as an industrial subcontractor for the young London start-up Humanoid (heise) : several hundred units targeted for 2027, a few dozen people dedicated. The machine assembled in Bühl is called Gamma and succeeds the HMND 01 prototypes, a Beta version being expected at the end of 2026 and mass production from 2028 (Denkstrom). Its geometry is not published anywhere that of its predecessor is, and it is not bipedal : the HMND 01 is a rolling platform with an omnidirectional base topped by a humanoid torso, a head and two arms, with a three-finger gripper in default configuration (The Robot Report). Humanoid does not sell its machines anyway : the model adopted is RaaS, the rental including fleet management, maintenance, 24-hour assistance and updates. Schaeffler, both investor and first customer of the robots produced in Bühl, commits to deploying a four-digit number of units on its global sites by 2032, starting with Herzogenaurach and Schweinfurt.
Same orientation elsewhere. Mobile humanoids have become the first source of revenue for Rainbow Robotics in the first half of 2026 (DigiTimes), on a half-year turnover of 21.38 billion won (Herald Corp). At Galbot, the machine working on a CATL production line is the S1, a heavy-load two-armed robot with a payload of 50 kilograms (CnEVPost) — not the first in-house biped, the ET1 ‘ Xingzai ’, unveiled on 20 August at the World Robot Conference (ITHome) then seen the day after tomorrow, according to the sole BigGo, in a tennis match against Zheng Jie, service over 100 km/h and retrieved without assistance after a fall, in autonomous decision-making. Remarkable demonstration, without a purchase order behind it.
There remains Xiaomi, the only biped of the week on a real line. After about four months of training in the group’s automobile factory, the success rate at the self-tapping nut placement station went from 90.2 % to 98 %, and the robot held three hours of continuous operation at a rate of 76 seconds per piece (36Kr). What legs bring to a screwing station remains to be demonstrated : no source describes the posture or the movements of the machine. The figure, however, turns as soon as it is related to the rate : 98 % at 76 seconds, that is one miss every fifty pieces, i.e. about one hour of production, and three continuous hours only make 142 cycles, a sample too short to conclude on a complete team. Xiaomi’s robotic activity remains at the prototype and internal validation stage, without its own external turnover (eWeek).
The producers, for their part, promise the opposite : XPeng targets the mass production of its humanoid IRON by the end of 2026, at more than 1 000 units per month (CnEVPost), and Unitree claims more than 5 000 shipments on the 2025 fiscal year (Yahoo Finance). Producer ambitions, not buyer commitments : a shipped volume says nothing about the format an industrialist chooses when signing.
Three deadlines will make the thesis verifiable. The first is close : the large-scale purchase wave of State Grid falls in the third quarter, and we will know if the line of 500 humanoids at 5 million yuan holds in firm order or deflates in favour of quadrupeds. End of 2026, XPeng will have to show a real monthly capacity rather than a target in August 2027, Bosch will have to get Gammas out of Bühl in series, for a client that has given itself until 2032 to deploy them. Until then, the only question to ask before each new record is that of the coefficient : 1.0 or 0.5.