Xiaomi humanoid robot runs on a live automotive assembly line for three consecutive hours

A humanoid robot has been demonstrated working fully autonomously on a live automotive assembly line for three consecutive hours.

That kind of sustained run puts human-shaped machines up against the tight timing and accuracy requirements typical of today’s car production.

Inside the car plant

At Xiaomi’s electric vehicle factory, the robot was assigned to a self-tapping nut installation station in the die-casting workshop, where alignment and tempo allow very little tolerance.

In a company post, Xiaomi recorded the trial and stated the robot achieved a 76-second cycle time with a 90.2% success rate, operating without any human intervention.

On that basis, the machine maintained the line’s quickest required cadence while fitting fasteners on both sides of the fixture.

However, simply reaching the necessary speed and baseline accuracy is only the initial hurdle, and the mechanical and control load behind each placement still warrants closer scrutiny.

Fasteners under pressure

For every cycle, the robot had to collect a nut from an automatic feeder and position it on a locating tool.

The station’s slide conveyor then advanced components into place, after which the robot tightened fasteners on a large floor section.

Because this was a live workstation rather than a staged showcase, the robot had to keep up with the line while maintaining correct tool and part alignment.

Falling out of step costs more than time: the next car body arrives regardless of whether the robot has finished its work.

Where mistakes happen

The main difficulty came from alignment: the robot needed to set each nut onto a pin without cross-threading.

A splined groove inside the nut had to engage cleanly, but each pickup subtly altered the approach angle.

Nearby metal also introduced magnetic attraction that could pull at the nut during contact, making tiny shifts significant.

Even with capable software, an awkward part to grasp can turn a minor misalignment into a jam that stops the station.

AI replaces scripted movements

To cut down on hand-written control rules, Xiaomi says it based the system on embodied intelligence-artificial intelligence that ties sensing, reasoning and motion together.

Instead of defining every trajectory in code, the robot learned from captured robot behaviour and then applied that learning back at the station.

At the centre is a 4.7-billion-parameter vision-language-action model – software that converts images and words into actions.

Even so, physical metal components rarely behave like tidy benchmarks, so the robot required more than cameras to remain dependable.

Sensing beyond cameras

During fastening, tactile feedback is crucial because vision alone cannot verify that a nut has seated flush against metal.

With force sensors added, the robot could feel contact load, helping to avoid blind tightening that might strip threads or bind components.

Input from joint proprioception-sensing limb position and movement-also helped keep the wrists stable when tools introduced sideways forces.

Improved sensing can boost repeatability, but it also means extra hardware and calibration that must endure heat, dust and vibration.

Coordinating arms and legs

Working inside the station required the humanoid to stay balanced while using its arms, rather than simply holding a fixed posture.

Whole-body control coordinated legs and arms together so the torso stayed steady during contact, instead of managing each limb in isolation.

Minor foot slips can grow into errors at the hands, particularly when the robot operates close to heavy moving equipment.

Greater stability opens the door to more demanding stations, yet every additional degree of freedom can make faults harder to anticipate.

Smoother robotic assembly

Line throughput also relies on fluid motion, as brief hesitations can cause tools to miss narrow timing windows.

To reduce stop-start movement, Xiaomi used asynchronous execution, allowing the system to plan upcoming actions while the robot continued moving.

By issuing short command batches, the robot could complete one sequence as software prepared the next, smoothing out transitions.

Continuous motion can recover valuable seconds, but only if sensing, computing and motors remain in lockstep amid factory noise.

Scaling across workstations

A single successful station does not make a robot ready for plant-wide deployment, so Xiaomi has begun testing additional routine assembly roles.

Alongside nut tightening, trials have included bin-picking and front badge installation-jobs that combine delicate grasping with rapid placement.

Hitting cycle-time targets leaves little room for the robot to pause and deliberate, so training needs to cover many infrequent edge cases.

High yield is still the metric that matters, because factories count good parts per hour rather than eye-catching demonstrations that require regular resets.

Rivals pick up speed

Rival efforts are accelerating as carmakers and technology firms try to move humanoids from rehearsed routines into real production environments.

“We do have some of the Tesla Optimus robots doing simple tasks in the factory,” said Elon Musk, Tesla’s chief executive officer.

If Xiaomi can improve consistency beyond these early results, manufacturers may begin purchasing humanoids in the same way they already buy industrial robot arms.

What factories gain

For practical factory adoption, the emphasis is shifting towards repeatable, data-driven motions supported by sensors and robust control, rather than one-off demonstration routines.

The next phase will probe whether such systems can operate through full shifts, cope with unusual parts, and fail safely when problems occur.