Easy for us, hard for machines

In 1988 the roboticist Hans Moravec pointed out something odd. It was fairly easy, he wrote, to get a computer to perform at adult level on intelligence tests or at checkers. It was hard, sometimes impossible, to give it the perception and movement skills of a one-year-old. This is now called Moravec’s paradox.

We rarely notice the hard part. A toddler picking up a sock uses skills learned so early that they feel like nothing at all. Chess feels hard to us, so we assume it must be hard for machines too. For a robot, the order is often reversed.

A mug and a T-shirt

Picture a coffee mug on a table. Once a robot knows where the mug is and which way it faces, it knows its whole shape. Push it and it slides, keeping its shape.

Now picture a T-shirt dropped on the same table. It can lie flat, fold over on itself, twist into a rope or bunch into a heap. A 2025 review of robotic cloth handling says soft objects break basic robotics assumptions, such as rigid shape and a simple way to describe what state the object is in. When a robot pushes cloth, the cloth moves and changes shape at the same time.

That creates two problems. The robot has to work out what shape the cloth is in right now, and parts of it are often hidden under other parts. Then every pull it makes produces a new shape it has to read again. The same review notes that many methods are built for one kind of fabric or garment and struggle to carry over to others.

What robots can do now

There has been real progress. In late 2024 the company Physical Intelligence described π0 (said “pi zero”), a robot model trained on data from many different robots. Its blog shows a robot unloading a dryer, bringing the clothes to a table and folding them into a stack. The company calls this task “exceptionally difficult for robots”, and says its robot can recover when someone interferes partway through.

The same company is plain about the limits. It calls π0 “only a small early step” and says general-purpose robot models are “still in their infancy”. Its videos include failed attempts as well as successes. Its research paper scores laundry folding in points for each step the robot completes, over 10 tries, and gives no single success rate.

So a robot folding laundry on video is real. A robot that folds any shirt, in any home, every time, with no person standing by, is something none of these sources claims.

In one sentence: skills that feel effortless to us are often the hardest for robots, and cloth is extra hard because its shape changes with every touch.