The Fascination With Robots Folding Clothes Explained

It seems like every week there’s a new video of a robot folding clothes. We’ve had some fantastic demos, like this semi-autonomous video from Weave Robotics on X.

That’s great, but Weave is far from the only company producing these kinds of videos. Figure 02 folds clothes. Figure 03 folds clothes. Physical Intelligence launched its flagship vision-language-action model, pi0, with a stunning video of a robot folding clothes after unloading a washing machine. You can see robots folding clothes live at robotics exhibitions. Even before all this, Google was showing folded clothes in its work, ALOHA was unleashed. 7X Tech is even considering selling robots to fold clothes!

And besides folding real clothes, there are other tasks similar to folding clothes, like Dyna’s towel folding, which leads to what is probably my best robot video of the year, demonstrating 18 hours of continuous towel folding. So why have all these robotic manipulation companies suddenly folded?

Reason 1: We basically couldn’t do this before

There is work going back over a decade that shows some amount of folded robotic clothing. But these demos were extremely fragile, extremely slow, and not even close to being production ready. Previous solutions existed (even learning-based solutions!), but they relied on precise camera calibration or carefully hand-crafted features, meaning that these clothing folding demonstrations typically only worked on a single robot, in a single environment, and may have only worked once, just enough for recording a demo video or paper submission.

With the help of a creatively patterned shirt, PR2 folded things up in 2014.Bosch/IEEE

Take a look at this PR2 folding laundry example from UC Berkeley from 2014. This robot actually uses a neural network policy. But this policy is very limited and fragile; it selects and places objects on the same green background, moves very slowly, and cannot handle a wide range of shirts. Making this work in practice would require larger models, pre-trained on web-scale data, and better, more general imitation learning techniques.

So, 10 years later, with the right demo data, many startups and research labs were able to implement clothing folding demonstrations; This is something we’ve seen from many hobbyists and startups, using broadly similar tools (like HuggingFace’s LeRobot), without intense specialization.

Reason 2: It looks great and people want it!

Many of us who work in robotics have that “north star” of a robot butler that can do all the tasks we don’t want to do. Mention folding clothes, and many, many people will tell you that they never want to fold clothes again and are willing to part with virtually any amount of money to achieve it.

This is also important for the companies involved. Companies like Figure and 1x have raised large sums of money on the idea that they would be able to automate many different jobs, but these companies increasingly seem to want to start at home.

Dyna Robotics can fold an indefinite number of napkins indefinitely.Dyna Robotics

And that’s part of the magic of these demos. Even though they are slow and imperfect, everyone can begin to imagine how this technology becomes what we all want: a robot that can exist in our home and alleviate all those daily annoyances that take up our time.

Reason 3: It avoids areas that robots are still bad at

These robot behaviors are produced by models trained via imitation learning. Modern imitation learning methods such as Diffusion Policy use techniques inspired by generative AI to produce complex and dexterous robot trajectories, based on examples of expert human behavior provided to them – and they often need many, many trajectories. Google’s ALOHA Unleashed is a great example, requiring around 6,000 demonstrations to learn, for example, how to tie a pair of shoelaces. For each of these demonstrations, a human piloted a pair of robotic arms while completing the task; all this data was then used to develop policy.

We must keep in mind what is difficult about these protests. Human demonstrations are never Perfectnor are they perfectly consistent; for example, two human protesters will never grasp the exact same part of an object with a precision of less than a millimeter. This is potentially a problem if you want to screw a cover onto a machine you’re building, but it’s not a problem at all for folding clothes, which is quite forgiving. This has two knock-on effects:

• It’s easier to gather the demos you need for folding clothes because you don’t need to throw out every practice demo that’s a millimeter over spec.
• You can use cheaper, less repeatable hardware to accomplish the same task, which is useful if you suddenly need a fleet of bots collecting thousands of demos, or if you’re a small team with limited funding!

For similar reasons, it’s great that with fabric folding you can fix your cameras in the right position. When you learn a new skill, you need training examples with space “coverage” of the environments you expect to see at deployment time. So the more control you have, the more efficient the learning process will be: the less data you need and the easier it will be to get a flashy demo. Keep this in mind when you see a robot folding objects on a regular table or with an extremely clean background; it’s not only a nice frame, it helps the robot a lot!

And because we’ve committed to collecting a ton of data (dozens of hours) to make this task work properly, mistakes will be made. So it’s very useful, if it’s easy to reset the task, that is, restoring it to a state from which you can retry the task. If something goes wrong, folding clothes is no big deal. Simply pick up the cloth, put it down and it’s ready to start again. This wouldn’t work if, for example, you were stacking glasses to store in a cupboard, because if you tip the stack or drop one on the floor, you’re in trouble.

Folding clothes also avoids violent contact with the environment. Once you apply a lot of pressure, things can break down, the task can become irresettable, and protests are often much harder to bring together because the forces are not as easily observable by politics. And each element of variation (like the amount of force you exert) will eventually require more data so that the model has “coverage” of the space in which it is intended to operate.

What to expect

Although we are currently seeing many clothing folding demonstrations, I still feel, overall, quite impressed by many of them. As mentioned above, Dyna was one of my favorite demos this year, mainly because longer-lasting robotic policies have been so rare until now. But they were able to demonstrate no-shot folding (that is, folding without additional training data) at several different conferences, including Actuate in San Francisco and the Conference on Robot Learning (CoRL) in Seoul. This is impressive and actually very rare in robotics, even today.

In the future, we should hope to see robots that can handle more difficult and dynamic interactions with their environment: moving faster, moving heavier objects, and climbing or handling adverse terrain while performing manipulative tasks.

But for now, remember that modern learning methods have their own strengths and weaknesses. It seems that, while not easy, folding clothes is the kind of task that really suits what our models can do right now. So expect to see a lot more.