The lights never dim on a tablet production line. In a factory in Nanchang, a row of slender machines kept working through the night, picking up unfinished devices, running them through tests, and sorting the good from the bad.
For years the public image of the humanoid (人形机器人) was a machine that could dance, bow, or wave on a stage. The harder question was always whether such a machine could hold a real job.
That question moved from theory to evidence this summer, when a cluster of general-purpose robots ran a continuous, unscripted shift inside a live electronics plant — and the record behind it is worth a careful look.
What actually happened on the line
The company behind the test is AGIBOT (智元机器人), one of the better-funded Chinese developers of embodied intelligence (具身智能) hardware. Across six days in late June, eight of its Genie G2 robots were placed inside Longcheer Technology's tablet manufacturing plant in Nanchang and set to work on a full quality-inspection section of the line.
This was not a single robot in a quiet lab. The machines worked the plant's own schedule, from 8 a.m. to 7 p.m., alongside human operators, moving material and absorbing the real rhythm of production. Eight robots handled the multimedia, audio, radio-frequency and coupling checks that a tablet must pass before it ships.
According to provincial industry reporting on the run, the robots completed:
- 64,828 individual work operations
- 17,625 finished products processed
- a reported task success rate of 99.99 percent
A separate milestone landed on the final day. AGIBOT said its 15,000th general-purpose embodied robot rolled off the production line and was delivered to Longcheer for use on the shop floor.
Why a boring shift beats a flashy demo
A dance routine proves a robot can balance. A six-day live shift proves something harder: that the machine can tolerate shifting light, moving people, rearranged materials, and the small chaos of a real plant without falling apart.
According to Yao Maoqing, senior vice-president at AGIBOT (智元机器人), the key question for humanoid robotics is no longer only what a robot can demonstrate, but whether it can be deployed, integrated, and create value in real operating environments. The six-day broadcast was meant as an open answer to exactly that question.
The shift in emphasis matters for the whole field. As embodied intelligence moves from research showcase to factory, the metric customers care about changes from "how impressive was that move" to "did the line keep running, and did it save money." That is a higher bar, and most companies are not yet close to clearing it.
The quiet backstory: a data flywheel
One reason AGIBOT pushed the public test is less about marketing and more about data. Every hour a robot spends in a real plant generates examples of what worked and what failed — precisely the material needed to train the next version of its control software.
This is the part global readers should watch. China's huge manufacturing base is not just a market for robots; it is a giant training ground. More deployed machines produce more behaviour data, better models, and in turn more useful machines. The loop is slow, but it is real, and it is hard to replicate outside a country with this density of factories.
The company says it has built dedicated data-collection facilities and treats high-quality real-world behaviour as the bottleneck that matters most. Whether that claim holds up will depend on whether the data actually translates into robots that need less hand-holding next year.
The global comparison
Western humanoid programmes have taken a similar turn. A US company demonstrated a robot sorting packages in a warehouse, and a major automaker is training its own machines on its own lines. The race is no longer about who can show the most human-like motion. It is about who can put a machine on a payroll-equivalent basis and keep it there. China's edge is the sheer number of factories willing to let robots try.
How this differs from a year ago
A shorter test earlier in the year used four robots for a single eight-hour broadcast. The June version scaled to eight machines, a full inspection section, and six consecutive days. That progression — from one workstation to a whole segment, from hours to days — is the kind of evidence that separates a press event from a deployment.
Other Chinese makers are pursuing the same path with different emphases. Some are targeting warehouses, others automotive lines, others retail floors. The common thread is a move away from "look what it can do" toward "here is the task it completed, for how long, and at what cost."
What is still unproven
The Nanchang test covered one type of work — tablet inspection — on a single production section. It did not show the robots swapping freely between unrelated tasks, nor did it publish the cost per unit of output or the maintenance burden when a joint wears out. A 99.99 percent success rate on a controlled station is not the same as proving a robot is cheaper than a person across a full year.
Honest limitations
The figures above come from the company and from provincial industry reporting, not from an independent auditor, so the 99.99 percent rate should be read as a vendor claim rather than a certified measurement. The deployment is impressive but narrow: it covers one inspection workflow, not a whole factory. We do not yet have public data on failure modes, repair costs, or payback period. Human oversight was present throughout, so this is assisted labour rather than autonomous replacement. Broader deployment plans for the year remain projections, not delivered results.
What readers can do now
- Treat "humanoid worked a real shift" headlines with a simple test: ask which task, for how long, and whether the cost was published.
- If you buy from or invest in robotics suppliers, request deployment evidence tied to a specific workflow and a time span, not a single demo video.
- Follow whether AGIBOT and its peers release independent uptime or maintenance data — that, more than any launch, will signal when the technology is ready for scale.