Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Can't Fold a Fitted Sheet
Transcript
- Lucas: So I tried something this weekend. I folded a fitted sheet — by hand, like a normal person — and I got about halfway through, realized I'd created a sort of fabric origami monster, and gave up and stuffed it in the linen closet. Luna: Okay, that is a universal experience. But I suspect you're not bringing this up just to confess your laundry failures. Lucas: No. Because while I was wrestling with it, I thought — we've done eighteen episodes now on things robot arms can't do. Open a door. Peel a potato. Handle a wet bar of soap. But we've never talked about the fitted sheet. Luna: And there's a reason for that. It's not just a funny laundry problem — it's actually a really deep robotics problem. Lucas: Exactly. So today I want to make the case that the fitted sheet is a perfect test case for a whole class of manipulation challenges that robots are terrible at: deformable objects. Non-rigid, stretchy, floppy things. Luna: Right. And we've danced around this before — the wet bar of soap episode, the overripe tomato — but the sheet is different. It's not just soft, it has structure. Lucas: It has corners. Elastic. And a continuous surface that can take on an infinite number of shapes. That's the part that breaks the robot's brain. Luna: So let's start with the basics. What makes a fitted sheet geometrically hard? Lucas: Okay, so imagine you have a rectangular piece of cloth with four elasticized corners. The robot needs to pick it up from a crumpled pile, identify each corner, match them — corner A to corner A, B to B — and then fold them together in a specific sequence. But the sheet can be in any configuration when it starts. Luna: And the corners aren't labeled. They all look the same. And they're buried under folds. Lucas: Right. So step one: perception. The robot has to look at a pile of fabric and figure out which pixel belongs to which corner. That's a segmentation problem. And fabric has very little visual texture — it's all one color, usually, so there are no landmarks. Luna: There's been work on this. I remember a paper from MIT CSAIL a couple of years ago — they used a special high-friction gripper and a lot of data to teach a robot to fold towels. But towels are simpler: they're rectangular, no elastic. Lucas: Right. Towels are rigid in a way — they have straight edges and right angles. A fitted sheet has curved edges when stretched, and the elastic creates tension that changes as you pull it. So the physics is non-linear. Luna: And that's before we even get to the folding itself. The robot has to plan a sequence of grasps and folds that don't just move the sheet, but also manage the state of the fabric — is it bunched? Is it twisted? Lucas: There's a term for this: state estimation under occlusion. The robot can't see the whole sheet at once because parts of it are hidden by other parts. And it can't predict how the fabric will behave when it lifts it, because the physics — friction, bending, stretching — is too complex to simulate accurately in real time. Luna: So brute-force simulation won't work. What about learning? DeepMind published a paper in 2025 — I think it was called 'Learned Folding Primitives for Deformable Objects' — where they trained a policy on millions of simulated folds. Lucas: Yeah, I read that one. They used a simulator called PlasticineLab — it's a differentiable physics engine that lets gradients flow through the material properties. So the network can learn how to fold by trial and error, but in simulation it can run a million trials in a day. Luna: And what were the results? Could it fold a fitted sheet? Lucas: In simulation, yes — with a simplified sheet that had only two elastic corners, and with full state information. But when they transferred it to a real robot, the success rate dropped to about thirty percent. The sim to real gap was huge. Luna: Because the real fabric wrinkles differently. The friction changes. The lighting creates shadows that confuse the camera. Lucas: Exactly. And the gripper — most robot grippers are designed for rigid objects. They pinch or suction. But fabric needs to be grasped in a way that doesn't create new wrinkles or slip. There's a whole subfield of soft grippers for fabric, but they're not robust enough for production. Luna: So where does that leave us? Are there any commercial robots that can fold laundry? Lucas: There are a few. There's a company called Foldimate — they make a consumer device that folds shirts and towels. But it's not a robot arm, it's a specialized machine — you feed the shirt in flat, and it uses rollers and plates. It can't handle a crumpled pile. Luna: And there's a Japanese company, Seven Dreamers, that tried to make a laundry-folding robot called Laundroid. They spent a billion yen, went bankrupt in 2019. The robot took about ten minutes per item and still got confused by collars. Lucas: That's the thing — the fitted sheet is the final boss of laundry. If a robot could fold a fitted sheet reliably, it could probably handle any piece of clothing. And that would unlock not just consumer laundry robots, but industrial applications — like automated textile recycling, where you need to sort and fold garments before shredding. Luna: Or surgical robotics — in the OR, surgeons need to manipulate drapes and gauze, which are deformable. If a robot could handle a fitted sheet, it could handle a surgical drape. Lucas: Exactly. So the problem is harder than it looks, but the payoff is real. And there's been progress. In 2024, a team at UC Berkeley published a paper on 'Folding in the Wild' — they used a dual-arm robot with two grippers and a learned policy that could fold a fitted sheet from a randomized pile with about sixty percent success. Luna: Sixty percent is not production-ready, but it's light-years ahead of where we were five years ago. Lucas: And they did it by breaking the problem down. First, the robot performs a 'flattening' motion — it uses one arm to smooth the sheet on a table. Then it uses a corner-detection network — trained on synthetic data — to identify the four corners. Then it uses a separate policy to pick up each corner and fold. Luna: That modular approach makes sense. You don't try to learn the whole task end to end. You learn sub-skills and chain them. Lucas: And that's the key insight. The fitted sheet problem isn't just about better grippers or better simulation. It's about task decomposition. Humans don't fold a fitted sheet by thinking about every wrinkle — we have a mental script: find a corner, tuck it into the opposite corner, repeat, then fold the resulting shape. The robot needs that same script, but it has to learn it without being told. Luna: So the real challenge is hierarchical planning. The robot has to decide which sub-task to do next based on the current state of the sheet. And that state is hard to measure. Lucas: Right. And that's where I think the next breakthrough will come — not from hardware, but from better state estimation. If you can give the robot a real-time 3D model of the sheet's geometry — even a coarse one — then you can apply classical planning algorithms on top of learned primitives. Luna: There's some work on using radio-frequency tags sewn into fabric to track points — like little QR codes for cloth. But that's not scalable. Lucas: No, but it's a research direction. And honestly, that's what makes this field so interesting — it forces you to think about perception, physics, planning, and hardware all at once. The fitted sheet exposes the limits of every subsystem. Luna: And it's also a great reminder that the hardest problems in robotics aren't the ones that look hard — like walking on two legs — they're the ones that look easy, like folding a sheet. Lucas: Yeah. And speaking of things that look easy but are actually hard — keeping a podcast going without ads or sponsorship is also a surprisingly delicate manipulation problem. It takes a lot of behind the scenes effort to research these episodes, and listener support is really what keeps it rolling. Luna: Absolutely. And we know our listeners get value from these deep dives. So if today's conversation gave you something useful, a couple of dollars a month genuinely makes a difference — it's what keeps the show independent and ad-free. Lucas: You can find us at buy me a coffee dot com slash fexingo. That's buy me a coffee dot com slash fexingo. No pressure, just if you've gotten something out of it. Luna: And now back to the sheet. So we've established it's hard. But let's zoom out — why do we care? Is this just about laundry? Lucas: It's not. Because deformable object manipulation is one of the last frontiers in robotics. We can assemble iPhones, we can perform surgery, we can drive cars. But we can't reliably handle a piece of fabric. And that limits what robots can do in homes, in hospitals, in textile manufacturing. Luna: So solving the fitted sheet isn't just about saving us from folding our own laundry — it's about unlocking a whole category of tasks that are currently out of reach. Lucas: Exactly. And the good news is that the research community is taking it seriously. There are workshops at ICRA and RSS specifically on deformable object manipulation. The number of papers has doubled in the last three years. Luna: So maybe in five years, we'll have a robot that can fold a fitted sheet. And then we can retire from this podcast. Lucas: Or we'll find something even harder to tackle. I'm thinking... why robots can't fold a fitted sheet while also making coffee. Luna: Let's not give ourselves ideas. Thanks for listening, everyone.