Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Fold a Towel
Transcript
- Lucas: So, we've spent fifty-six episodes watching robot arms fail at tasks that any human finds trivial. Gripping a raw egg. Opening a door. Threading a needle. Luna: And today we're adding another one to the list. Lucas: Today: folding a towel. Not a surgical gown, not a fitted sheet — a simple, rectangular, terry-cloth bath towel. Luna: And the robot still can't do it. Lucas: It can't. And the reason tells us something fundamental about where robotics is stuck in mid-2026. So let's start with why this is hard. Luna: Because cloth is not rigid. Lucas: Exactly. Every episode we've done on manipulation — the egg, the tomato, the zipper — those objects are rigid or semi-rigid. A towel is what roboticists call a deformable object. Its shape isn't fixed. It bends, it wrinkles, it drapes. And the number of possible configurations is effectively infinite. Luna: So the robot's vision system can't just look at it and know where the corners are. Lucas: It can guess, but not reliably. In fact, there's a well-known paper from 2018 at MIT — they used a two-armed robot to try to fold towels from a laundry basket. The robot had to pick up the towel, identify the corners, and then bring them together. It succeeded about 50 percent of the time. Luna: Fifty percent? That's basically flipping a coin. Lucas: And that was in a lab, with good lighting, with a single towel. In the real world — a messy laundry pile, different fabrics, different sizes — the success rate drops to single digits. Luna: So what's the core problem? Is it the gripper? The vision? The planning? Lucas: All three, but let's start with vision. A robot sees the world through cameras and depth sensors. For a rigid object — say, a coffee cup — the vision system can build a 3D model, track its pose, and plan a grasp. For a towel, the shape changes every time you touch it. So the model is wrong as soon as it's built. Luna: So it's like trying to catch a moving target that also changes shape. Lucas: Exactly. And the gripper problem is just as bad. Most industrial grippers are designed for rigid objects — parallel jaws, suction cups, magnets. A towel is soft, porous, and floppy. Suction cups often leak air. Jaw grippers crush the fabric. There's some work with electrostatic grippers — they use a voltage to make the fabric stick — but that's still experimental. Luna: And then there's the planning. Even if you could see the towel and grip it, you still have to decide how to move to fold it. Lucas: Right. Folding a towel requires a sequence of actions: pick up, bring two corners together, then bring the other two, then set down. But the physics of cloth is nonlinear. A small error in the first fold propagates into a huge misalignment by the third fold. Luna: And humans do this without thinking. We use tactile feedback — we feel when the corners align. Lucas: That's a huge part of it. Human hands have thousands of tactile sensors. Robot hands have, at best, a few dozen. So the robot is operating almost blind in terms of touch. It's relying on vision that's already lagging behind. Luna: So what's the state of the art in 2026? Has anyone solved it? Lucas: Not solved, but there's progress. A group at the University of California, Berkeley, published a paper last year using something called a 'garment graph' — a data structure that tracks the key points of a piece of clothing as it moves. Their robot could fold a T-shirt with about 80 percent success in a controlled setting. Luna: But not a towel? Lucas: Towel is harder. A T-shirt has structure — seams, collars, sleeves — that give the vision system anchor points. A towel is just a rectangle of fabric. So the garment graph approach struggles. Luna: So what about the MIT work you mentioned earlier? Did they improve? Lucas: They did. In 2024, the same lab — now led by a researcher named Daniela Rus — published a new method using differentiable cloth simulation. Basically, they run a physics simulation in parallel with the real-world motion, and use the simulation to predict how the cloth will behave. Then they adjust the robot's actions accordingly. Luna: So it's like a digital twin for the towel. Lucas: Exactly. And that simulation approach improved fold success to about 60 percent. But the simulation runs slower than real time — so the robot has to move slowly. A human can fold a towel in about four seconds. The MIT robot takes about forty seconds per fold. Luna: So it's not ready for commercial use. Lucas: Not yet. And that's a big deal, because there's a huge market here. The global laundry automation market was estimated at about two point three billion dollars in 2025, and it's projected to grow to three point eight billion by 2030. Hotels, hospitals, industrial laundries — they all want to automate folding. Luna: And they've been waiting for a while. Lucas: Decades. There were attempts in the 1990s. A company called FoldiMate tried to build a consumer folding machine — they showed a prototype at CES in 2019, but they never shipped. They went out of business in 2021. Luna: So the problem is harder than VCs thought. Lucas: Much harder. But there's a new wave of startups. One I'm watching is a company called Garment Robotics. They're based in Shenzhen, and they're focusing on industrial laundry — the kind that processes thousands of uniforms per day. They use a combination of suction grippers and computer vision trained on millions of images of crumpled fabric. Luna: And how are they doing? Lucas: They claim ninety percent success on standard work shirts. But towels? They won't even touch them. Luna: Because towels are the hardest. Lucas: They are. And that's why I think towel folding is a kind of benchmark for the field. If a robot can fold a towel reliably, it means it can handle any deformable object. And that would unlock robotics in a whole range of industries — not just laundry, but also food handling, medical textiles, even soft packaging. Luna: So it's a moonshot. Lucas: It is. And we're seeing more investment. The Defense Advanced Research Projects Agency — DARPA — has a program called 'Deformable Manipulation' that started last year. They're funding five teams to work on cloth folding and cable routing. So the smart money is betting that we'll see a breakthrough in the next three to five years. Luna: If today's episode gave you something to think about, and you'd like to see more deep dives like this, the way we keep the show ad-free is through listener support. You can find us at buy me a coffee dot com slash fexingo. Anything helps, and it really does make a difference. Lucas: Yeah, we appreciate that. It helps us keep digging into these topics without commercial pressure. So, back to the towel. One more piece of research I want to mention — a team at Stanford recently showed that if you give a robot arm a specialized 'cloth gripper' that uses controlled air jets to flatten fabric, you can improve fold accuracy significantly. Luna: So they're using airflow to pre-smooth the fabric? Lucas: Yes. The gripper has tiny air nozzles that blow the fabric flat against a table before the robot attempts a fold. That reduces the deformable-object problem to a rigid-object problem — at least for the first fold. Luna: Smart. But it adds complexity and cost. Lucas: It does. And it only works on a flat surface. So if you're folding a towel straight out of a laundry basket, you'd need an extra step to lay it flat first. But it's a sign that researchers are getting creative. Luna: So where do you think we'll be in five years? Lucas: I think we'll have a commercial robot that can fold standard bath towels with high reliability — maybe ninety-five percent or better — but only in controlled environments like hotels or hospitals. Home use, where every towel is different and the lighting varies, that's probably ten years away. Luna: So for now, we're still folding our own towels. Lucas: We are. And honestly? It's not that bad. It takes about four seconds. But for the robotics industry, those four seconds represent one of the hardest unsolved problems in manipulation.