For five months, the same criticism has recurred regarding the humanoid sector: nowhere is there verifiable proof of work, only choreographies, in-house rankings and view counters. The week of 14 to 20 April 2026 saw the two halves of the sector unfold almost on the same day — and, for the first time, one of them provided the missing proof. On one side, the spectacle reached its point of self-parody. On the other, a production line ran for eight hours, live, under the public eye. Between the two, a figure that does not lie: continuous running time.
Beijing, 19 April: the record decided after the fact
At the 2026 Beijing E-Town Half Marathon, the Lightning (闪电) robot from the Qitian Dasheng team, developed by Honor, won with a net time of 50 minutes and 26 seconds in autonomous navigation. On paper, the performance crushes the human: the men’s world record for the half marathon, 57 minutes and 20 seconds, set by Ugandan Jacob Kiplimo in Lisbon in March 2026, is beaten by nearly seven minutes.
The detail that devours the feat lies in a coefficient. An Honor robot, remotely piloted by the Jueying Chitu team, physically crossed the line first, in 48 minutes and 19 seconds. But the lack of autonomous navigation earned it a 20% penalty — a 1.2 coefficient applied to its net time — which pushes it behind the official winner. In other words, it was not the race that separated the machines, but the rules, and these rules decide after the fact what counts. The victory is first and foremost an accounting entry.
The context further relativises the "record". Only 38% of the robots entered were running in fully autonomous navigation; all others were remote-controlled by human operators. And the big winner is not a roboticist: Honor — Shenzhen Rongyu Intelligent Technology — is first known as a manufacturer of smartphones and consumer electronics. The sector celebrates a speed, but the majority of the peloton was not deciding its own trajectories.
Nanchang, 14 April: eight hours, without cuts or re-editing
Five days earlier, at the Longqi Technology tablet factory in Nanchang in Jiangxi, the Elf G2 robot from Zhiyuan Robotics sustained an uninterrupted public live stream of eight hours on a 3C production line under real conditions. Precision counts here as much as duration: it was neither a demonstration nor a rehearsal. The Elf G2s were running on the real line during the broadcast, without interruption or editing — exactly what the sector had been demanding for months.
The figures hold up. The robot processed 310 tablets per hour, on a cycle of 18 to 20 seconds per operation, equivalent to the workload of two sequential human posts. Over the eight hours, the overall success rate remained above 99.5%, with no major anomaly reported. Where the marathon measures a speed peak, the factory measures endurance — and it is this endurance, verifiable continuously by a third party, that constitutes the new metric.
Zhiyuan does not hide it: the company qualifies the episode as the world’s first industrial-scale embodied AI on a precision 3C line, and sees it as the entry of the technology into normal operation. The announced trajectory extends the gesture: a deployment of 100 Elf G2s at Longqi Technology by the end of the third quarter of 2026, with an extension aimed at automotive, semiconductors and energy. The robot, the press summarises, has moved from artist to worker.
The hard figure the week hides under the folklore
On 15 April, Roland Berger sets out the economic corollary of this shift. According to its report, a humanoid robot could be operated at an operational cost of approximately 2 dollars per hour, making its deployment competitive even in high-wage countries. In the consultancy’s reference scenario, the OEM market for humanoids would reach 300 billion dollars in 2035, and 2,000 billion in 2050.
The same report plants the thorn. Hardware is currently three to five years ahead of software; supply chains remain immature and regulatory frameworks are fragmented. It is precisely this hardware/software fracture that the week stages: capable bodies, lagging brains.
π0.7, or the embodied fracture
On 16 April, Physical Intelligence provides the clearest illustration with its model π0.7. Generalisation is impressive, but the model cannot execute a complex multi-step mission from a single high-level command — preparing a toast in full autonomy, for example, remains out of its reach without step-by-step human guidance. The most instructive admission comes from the team itself: it is impossible to determine with certainty whether the model is truly solving a novel task by generalisation or recalling very similar training data.
The gap is then clearly readable. On a 3C line, the task is repetitive, framed, with no open decisional branch: the software maturity missing from π0.7 is not required there, and endurance is sufficient to make the proof. Industrial deployment does not wait for the software to catch up with the hardware; it chooses the tasks that the hardware already knows how to handle.
Hanover: the West catches up on substance, not on show
On the European side, the week confirms a repositioning on the same ground — that of industrial substance. At Hannover Messe 2026, a dozen or so manufacturers presented humanoids, the first time that industrial humanoid robotics occupied a central place in this trade fair. The most telling signal is not a dance, it is a component: Schaeffler won the Hermes Award 2026 for its platform of integrated actuators designed for humanoid joints — knowing that actuators represent, according to the industrialist, approximately 50% of the total cost of a robot. Value is played out in the mechanics, not in the choreography.
The rest follows the same logic. Agile Robots, a German company, presented the Agile ONE as the first industrial humanoid manufactured in Germany. And the BMW pilot in Leipzig — the first deployment of a humanoid in automotive production in Europe, after the Figure 02 tests conducted by BMW in Spartanburg in the United States — is not a break but a continuation, the next step in an already engaged programme. The event of the week is not there.
The right question has changed
The marathon and the factory took place within the same seven days, and it is this telescoping that makes the news. As long as the sector judged itself by view counters or the most contested ranking, the question was: who is number one? Now, a metric of substance exists — the hour of continuous operation under the eye of a third party, and its hourly cost — and it shifts the question. No longer who runs the fastest, but: how many hours, at what cost, and has it really been seen running? The week of 14 to 20 April was the one where the robot stopped dancing to start clocking in.