Xiaomi disclosed on November 3 that its humanoid robot completed an eight-hour shift at a factory sequencing station for the first time, correcting its own errors and syncing with the production line's control systems throughout the day. In a second trial, a home robot agent built on the company's Xiaomi-Robotics-1 model carried out more than 50 tasks across three rooms in a single continuous session lasting over 30 minutes. Xiaomi shared both results during the 2026 World Internet Conference, according to PANews.
The factory robot worked a center console side-cover sequencing station, the kind of repetitive parts-handling job normally staffed by a human line worker or a fixed-arm machine. Xiaomi says the robot's own anomaly correction kept the line running without a person stepping in to fix jams or misreads.
What makes an 8-hour shift a hard problem?
Most humanoid robot demos run for minutes, not hours, because errors compound. A small misgrip or a part that lands slightly off-angle can cascade into a stalled line if the robot cannot recognize and fix the problem itself. Running for a full shift means the robot has to handle hundreds of small variations in parts, lighting and timing without a human resetting it between attempts.
Xiaomi's robot is built on Xiaomi-Robotics-1, a model pretrained on 100,000 hours of real-world manipulation data captured through a portable recording rig, then fine-tuned on a smaller set of robot-specific trajectories. That two-stage approach, cheap human demonstration data first, targeted robot practice second, is the same scaling recipe labs elsewhere have used to push general-purpose manipulation past narrow, single-task automation. Xiaomi says the model topped the RoboCasa365 and RoboDojo benchmarks in July, two standard tests for how well a robot policy generalizes to tasks it was not explicitly trained on.
Why run a home trial at the same time?
The home trial, at what Xiaomi calls its youth apartments, tested something different from the factory run: not raw endurance, but the ability to string together many small, varied chores, more than 50 of them, across three separate rooms without losing track of the overall goal. That kind of long-horizon planning is widely seen as harder for current robot models than repeating one motion for hours, because the robot has to decide what to do next rather than just execute a fixed sequence.
Why does this matter beyond Xiaomi?
Xiaomi is one of several Chinese manufacturers, alongside AgiBot, Unitree and UBTech, racing to put humanoid robots into real factory and household settings rather than lab demos. An eight-hour unsupervised shift, if independently verified, would be a meaningful marker, since most public humanoid trials to date have been measured in single-digit hours at most, often with engineers standing by to intervene.
The disclosure also fits a broader pattern among Chinese robot makers of investing heavily in the software and data side of humanoid robots, sometimes called the "brain," rather than competing purely on hardware specs. Xiaomi's decision to publish benchmark rankings alongside the factory result suggests the company is positioning Xiaomi-Robotics-1 as a model it may offer more broadly, not just a one-off demo built for a single product line.