Eka's robots sort chicken nuggets and screw in light bulbs, as a ChatGPT moment for physical AI draws near

Eka has built robots that screw in light bulbs, sort chicken nuggets and move fragile objects, with movements that are eerily close to human ones. Wired asks whether this dexterity reflects real physical intelligence or a well-trained imitation of motion.

What Eka's robots can do

Eka's designs differ from standard industrial manipulators, which for decades worked in the auto industry along strictly predefined trajectories. Eka's grippers are not just "claws" but multi-jointed hands with tactile sensors in the fingertips. Instead of hard-coded trajectories, the systems are trained on video datasets of human labor and fine-tuned in simulation. This lets them adapt to objects of different shapes and weights without preset coordinates.

Here is what the same robot does in real-world conditions:

• Sorts chicken nuggets on a conveyor by shape and position
• Screws light bulbs into sockets at various angles
• Moves fragile eggs without damaging the shell
• Handles unfamiliar objects without point-by-point recalibration
• Works in an unstructured environment, where earlier robots required perfect order

The key difference from the previous generation is the ability to cope with variability. Not "perform movement A at point B," but "figure out this object the way a human would."

A ChatGPT moment or a skilled imitation

Wired raises an uncomfortable question: how far do these robots' movements reflect a real understanding of space, rather than an accurate reproduction of patterns from training data? For language models, the "parrot" problem, a model that imitates understanding without actually understanding anything, has never been fully resolved. With physical robots the question is sharper: an error on a production line costs more than a hallucination in text.

Eka and its competitors, Apptronik, Figure AI, Physical Intelligence (π) and 1X Technologies, are betting that foundation models for movement are already good enough for commercial use. But researchers disagree. Some believe the line between "physical intelligence" and "very good motion statistics" has no practical significance if the robot performs the task reliably. Others warn that imitation breaks down precisely in edge cases.

"The question is not whether a robot can screw in a light bulb.

The question is what it will do when the bulb turns out to have a different diameter," is how one robotics researcher frames the logic.

Why the threshold is here now

Cheaper actuators, affordable depth cameras, large datasets of human movement and a transformer architecture for physical AI all converged at once. The distance between a lab prototype and a working robot has shrunk faster than the industry expected. Eka is already testing its systems on real production lines in the food industry.

The parallel with language models is direct: ChatGPT launched in November 2022, and over the next 18 months the industry rethought what can be automated at all. Physical robots may be standing at a similar threshold. If so, decisions about which manual tasks to automate will be made not in 10 years but in 2-3.

What this means

If Eka and its competitors are right, the next few years will be a turning point for manual labor in manufacturing and logistics. The question is no longer "will robots replace human hands" but "in which tasks will it happen first."