Eight hours is not a demo number, it is a shift number. On 3 November Xiaomi posted video of two full-size humanoid robots working the centre-console side-cover sorting station in its own EV factory's final assembly shop, and they did it for 8 hours and 5 minutes without a human stepping in.

The robots pulled moulded interior panels out of bins, turned them the right way up and slotted them into the correct rack position for whichever car was coming down the line. That is the most credible factory humanoid result out of China this year. It is also still a demonstration, because the number that decides whether it is a product is the one Xiaomi did not publish: the success rate over those eight hours.

The progression is public, and it is not hype

Milestone Station Reported result
March 2026 Self-tapping nut loading 3 hours continuous, 90.2% bilateral install success, met the 76-second line takt
July 2026 Nut loading plus two new stations Success rate up to 98%; side-cover sorting and bin folding/return added
3 Nov 2026 Centre-console side-cover sorting 8 hours 5 minutes, met the station's full daily output

The first two numbers came from Beijing Daily's reporting and the 90.2% to 98% climb on the nut station is a real engineering figure. Publishing it is rarer than it should be in this industry, so credit where it is due.

The announcement landed during the World Internet Conference in Wuzhen, where Xiaomi took the conference's Outstanding Contribution Award and the robot team's embodied foundation model work took the Leading Technology Award. That is the context: this was presented as a research milestone in front of a tech crowd, not a product launch.

The job is harder than "put the part on the shelf"

Side-cover sorting is an SPS supply station in auto final-assembly logistics. The panels are large, flexible, irregularly shaped and arrive in random orientations, so the robot cannot rely on a fixed pick pose. It has to reorient each part in the gripper and match the rack slot to the sequence of vehicle models on the line, at takt. That demands dual-arm coordination, dexterous handling, force-controlled compliant motion and visual anti-interference at once. 澎湃新闻's write-up of the video makes the same point: the interesting claim is not that a robot moved a plastic panel, it is that a robot did the part of the job that has resisted fixed-path automation.

Electrek's complaint that Xiaomi withheld the 8-hour numbers is fair. The power question it raised does have an answer, though, and it comes from Chinese coverage of the video rather than from Xiaomi's own post: the robots charged twice during the shift, after 3 hours 2 minutes and again after 6 hours 3 minutes of cumulative work. So this is charging rather than one eight-hour cell, and the 8 hours 5 minutes is robot time with two charging stops in it. Honest, and also a cost line, because two charges per shift are two chances to fail.

The wheeled robot is the tell

Around 1 minute 37 seconds into Xiaomi's clip, a segment labelled "Dropped Material: Multi-Robot Collaboration" shows a humanoid upper body on a wheeled base rolling in to help the two legged robots. 澎湃's account adds the detail that matters: in the second half of the shift, one of the robots dropped a part on the floor and a wheeled robot came over to pick it up.

That sequence is the state of the art in one image. The legged robot does the hard work, drops things, and the cheap wheeled platform cleans up. Wheels cost a fraction of what legs cost, carry more, hold a bigger battery, do not topple when a panel hits the deck, and the autonomous mobile robots already shuttling bins around the same shop floor have been doing material movement for years. The valuable engineering is in the hands and the model, not in the legs.

Lei Jun's reply to netizens asking whether humanoids make sense in factories at all was the right answer: "This is a general-purpose humanoid robot. The hardware doesn't need to change, and it can adapt to more different stations and scenarios." True of the hardware. Every new station still needs its own training, its own validation and its own pile of edge cases, and that is where the money hides.

The harder problem is happening in the apartments

The same 3 November announcement put Xiaomi's open-source Xiaomi-Robotics-1 model into trial operation in Xiaomi's staff apartments, with an agent system built on it. It ran more than 30 minutes continuously across three rooms and completed more than 50 long-horizon tasks: tool preparation, sofa and desktop tidying, rubbish collection, sorting items back to where they belong, room inspection, a whole-home re-check and leaving the flat.

The model was released on 16 July with 100,000-plus hours of real-world manipulation trajectories collected with UMI grippers, trained in two stages (large-scale pretraining, then cross-embodiment post-training) and published as arXiv 2607.15330. Xiaomi says it is the first validation of scaling behaviour in an embodied foundation model.

A factory is a closed world: known parts, known slots, known takt. An apartment is an open world with unknown objects and unknown surfaces. If one model handles both, deploying a robot to a new station gets dramatically cheaper, because the training data stops having to be collected from scratch each time. That is the part of the announcement worth watching, and Electrek gave the flat one paragraph.

What to demand before "large-scale deployment"

Last December Lei Jun said Xiaomi would deploy humanoid robots in large numbers across its factories within five years. Here is the checklist to hold them to when that deadline arrives: a success rate per shift averaged over weeks rather than a best day; the number of human interventions and the mean time between them; takt performance at the 98% case, not the 90.2% one; hours of retooling and retraining per new station; and cost per station measured against the fixed robot doing the same sorting job, and against the AMRs already moving the bins.

For scale, Tesla's Optimus at the October 2024 "We, Robot" event was being teleoperated by humans, and Tesla now says it will not deliver Optimus to outside customers until late 2027. Figure 03 has been picking parts out of unsorted bins and delivering them to carts at BMW's Spartanburg plant, and Agibot's G2 has been running MMI testing at Longcheer's Nanchang plant on an 8am-to-7pm shift. Nobody has published a week of factory numbers, which is the industry's real problem, and Xiaomi is currently the only company publishing a success rate at all.

The verdict

This is genuine industrial progress and it deserves the credit: a robot that works a full production day on a flexible-parts station is not a stage demo. But an 8-hour video with no error rate, no part count and no cost is still a video, and the same factory floor was already full of wheeled robots doing logistics faster and cheaper.

Judge the next milestone the same way: not the hours, but the interventions per hundred hours. If Xiaomi publishes that number for the nut station and the side-cover station at the same time, the humanoid factory case is real. Until then, watch the wheeled one.

Sources