At Hyundai's plant, Atlas humanoid sorts roof racks without help

On the floor of Hyundai's auto plant, a humanoid robot now helps sort roof racks for the assembly line, and nobody guides its hands. Atlas, built by Boston Dynamics, stands 175 cm tall and weighs 90 kg. It works among more than 1,000 robots and nearly 1,500 people, who lift, stamp and weld parts in a plant that The Robot Report calls arguably the most advanced in the world. Atlas, the report says, "has left the lab and is doing real work."

To see how far that is from the start, look at 2021. Atlas was then a bulky robot powered by hydraulics, a system that moves limbs with pressurised fluid, and it could only run and jump. Its movements were written line by line by engineers. The current version has an all-electric body and runs its AI on NVIDIA chips, the same family of processors that powered the recent wave of generative AI. The change is less about the hardware than about how the robot learns.

The method is teleoperation: a person takes control. Machine learning specialist Kevin Bergerman, wearing a virtual reality headset, guided Atlas's hand and arm through each task until the robot could master it. Each guided run produces data, and that data trains the AI to repeat the task on its own. A suit with sensors can also record a person's movements, which serve as examples for the robot to imitate. Those examples are then fed into a simulation: over 4,000 digital Atlases were trained for six hours, facing slippery floors, bending and stiff joints. Once one robot has learned a skill, the report notes, the same skill can be loaded into all the others.

The limits are as clear as the progress. Atlas still can't do most of what a person does in a day: getting dressed or pouring a cup of coffee and carrying it through the house are beyond it. It managed crawling and cartwheels, but its duck walk was weaker. The robot dog Spot, introduced almost ten years ago, shows a narrower path that already works: it runs quality checks and industrial inspections at hundreds of facilities around the world. For a factory, the near-term gain is the repetitive, heavy or hazardous handling that already happens in a plant with automation. Whether that extends to general-purpose work, and when, the report does not say. The aim it describes is a robot that can be told what to do in plain language and understand how the physical world works. That is a goal, not yet a product.