Xiaomi's Humanoid Robots Just Hit 90% Accuracy on a Live EV Assembly Line — and Ran for Three Hours Straight

Humanoid robots in factories have been a staple of corporate keynotes and concept videos for years. Xiaomi just put numbers on what happens when they actually show up for a shift. The company tested its humanoid robots on a live electric vehicle assembly line in its die-casting workshop, where the machines achieved a 90 percent success rate installing self-tapping nut components and completed three consecutive hours of autonomous operation without human intervention.

What the Test Actually Involved

The deployment was not a controlled demo in a lab. Xiaomi placed its humanoid robots at an active workstation in the die-casting workshop of its EV manufacturing facility, tasking them with a specific, repetitive operation: installing self-tapping nuts on cast components. The robots had to locate the correct insertion points, orient the fasteners, and apply the right torque — a task that sounds simple but requires the kind of hand-eye coordination and force control that has historically been the domain of human workers or purpose-built automation cells.

Three hours of continuous autonomous operation is notable. Most humanoid robot demonstrations last minutes, carefully stage-managed to avoid failure modes. Running for three hours on a live production line means dealing with part variation, positional drift, and the ambient chaos of a real factory floor. A 90 percent success rate means roughly one in ten fasteners still required human correction — a number that would need to improve significantly before the robots could replace workers outright, but sufficient to demonstrate viability as an assistive system.

The Broader Humanoid Push

Xiaomi's test arrives in a market that is moving fast. A facility in Foshan, Guangdong province, is now producing one humanoid robot every 30 minutes, targeting 10,000 units annually across 24 digitalized assembly stages with 77 inspection checkpoints per unit. In Canada, seven humanoid robots are heading to Toyota Motor Manufacturing's Woodstock, Ontario RAV4 plant under a Robots-as-a-Service agreement — the first commercial humanoid deployment in Canadian automotive manufacturing.

Meanwhile, Hyundai announced at CES 2026 that it will begin deploying Boston Dynamics' Atlas humanoid in its Metaplant America facility in Savannah, Georgia, starting in 2028, with plans to scale to 30,000 humanoid units annually for factory automation. And in Europe, UK-based Humanoid completed a live trial inside an active automotive production facility in partnership with SAP and supplier Martur Fompak.

Why Humanoids Instead of Purpose-Built Arms

The obvious question remains: why deploy an expensive humanoid when a conventional robot arm can install fasteners all day without legs, a torso, or a head? The answer lies in flexibility and facility economics. Automotive plants are designed around human ergonomics — workstation heights, aisle widths, tool placement. Deploying a conventional automation cell at each station requires infrastructure modification: mounting plates, safety fencing, rerouted conveyors. A humanoid, at least in theory, slots into the existing workspace with minimal physical changes.

The theory is ahead of the economics. At current price points, humanoid robots are far more expensive per task than either human workers or dedicated automation. But if production volumes hit the targets that Chinese manufacturers are setting — 10,000 units a year from a single facility — costs will drop. And if the 90 percent success rate climbs toward 99 percent with improved training data and better manipulation algorithms, the total cost of ownership equation starts to shift.

What's Left to Prove

Three hours is not a shift. Ninety percent is not production-grade. Xiaomi knows this, and the test was positioned as a milestone rather than a finished product. But milestones matter. Two years ago, the idea of a humanoid robot operating autonomously for hours on a real production line was speculative. Today it's a data point — imperfect, but real. The gap between concept video and factory floor just got measurably smaller.