Learn

AI meets robotics

For decades, robots were precise but rigid. The same kind of AI behind chatbots is now teaching them to handle the messy world.

The old way

Traditional industrial robots follow programs written line by line. They weld the same seam or move the same part thousands of times a day, very accurately, inside cages. Change the part or the layout and an engineer has to reprogram them. They do not understand anything; they repeat.

Moravec's paradox

In the 1980s, roboticist Hans Moravec noticed something odd: tasks that are hard for humans, like advanced math, turned out to be easy for computers, while tasks a toddler does without thinking, like picking up an unfamiliar object or walking over clutter, were extremely hard. That paradox is the main reason software AI has raced ahead of robots.

What changed

Large models trained on internet-scale text and images learned a broad, loose understanding of objects, language, and cause and effect. Researchers then began training vision-language-action (VLA) models that take in camera images and a plain instruction, such as "put the red cup in the sink," and output robot movements directly. Google DeepMind's RT-2 work in 2023 was an early landmark; since then, Gemini Robotics, Figure's Helix, Physical Intelligence's models, and Nvidia's GR00T effort have pushed the same idea further.

The payoff is generalization: a robot that can attempt tasks and objects it was never explicitly programmed for.

Why humanoids

Factories, warehouses, and homes are built for human bodies: human-height shelves, stairs, door handles, tools. A humanoid form can, in principle, work in those spaces without rebuilding them. Humanoids also generate training data that maps naturally to video of people doing tasks. The tradeoff is cost and complexity; wheels and simple arms are still cheaper for many jobs.

What still holds robots back

Data is scarce: there is no internet of robot movements, so companies collect it through teleoperation, simulation, and fleets in the field. Reliability is unforgiving: a chatbot can be wrong 1% of the time, but a robot dropping 1% of parts is unusable. Hands remain hard, batteries limit shifts, and safety around people sets a high bar. As of 2026, most real deployments are narrow jobs like moving totes and parts.

For which robots are actually working, and where, see the tracker at sirobot.us.