AgiBot put four G2 humanoids on a tablet assembly line and live-streamed an eight-hour shift. The interesting part is not that they looked vaguely human while doing it. It’s the very specific, very factory-shaped numbers being waved around.
The news hook (and why it’s unusually concrete)
In Nanchang, a tablet manufacturing workshop ran a live-streamed shift with four humanoid robots doing final inspection and a tight pick-place-test loop. The reporting claims ~18–20 second cycles, ~310 units/hour, and success rates north of 99% (Xinhua via SCIO). Interesting Engineering repeats similar figures and adds the kind of detail companies love when they think they’re about to win procurement: “integrated in 36 hours”, “3,000 units per shift”, and “downtime loss below 4%”.
That is either (a) the first real “humanoid in a boring factory loop” case study with enough numbers to argue about, or (b) a beautifully lit pilot where the failures were asked to wait off-camera until the livestream ended. Possibly both.
What these numbers actually mean
Cycle time is real. It tells you the robot can hit a cadence that resembles wage math, not demo math.
Changeover time is even more real. The SCIO/Xinhua piece claims calibration in minutes and changeover/retraining in hours. If that holds outside a staged line, it’s a serious signal because factories change SKUs the way startups change priorities.
“Success rate” is the slipperiest metric in the bundle. 99.9% of what, exactly, and measured how? One cherry-picked station is not the same thing as a week of mixed products, dusty sensors, bumped fixtures, and operators who do not care about your company’s valuation narrative.
140 hours of continuous operation is a start, not a victory lap. The only number that turns this into deployment, rather than theatre, is uptime over months plus a clear definition of what counts as intervention.
The Droid Brief Take
Humanoid robotics keeps trying to sell you the future with backflips. China is currently trying to sell you the future with throughput.
If these factory figures are even directionally true, the real story is not “humanoids are here.” It’s that the industry is learning the only language factories speak: integration time, yield, downtime, and whether the line boss stops getting paged at 2am.
Also, note the quiet admission hiding inside the triumph: the hard parts are force control and tactile sensing. You can have all the embodied intelligence you want. If your robot can’t reliably feel what it’s doing, it’s just an expensive way to drop tablets.
What to Watch
Definition of “minimal human intervention”. Is there a human babysitter, a teleop shadow, or an exception handler who is doing heroic work off-screen?
Time window. Do these metrics hold after the novelty period, when the line gets messy and nobody is carefully curating conditions?
Scaling claim. Multiple sources cite expansion to 100 units by Q3 2026. Watch for: where, doing what tasks, under what service model (spares, maintenance, MTTR).
Sources
State Council Information Office (Xinhua) — “Chinese humanoid robots deployed on assembly lines for precision tasks”
Interesting Engineering — “Watch ‘world-first embodied AI’ humanoid robots take on manufacturing tasks in China”