Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Fold a Fitted Sheet
Transcript
- Lucas: So we've spent over a hundred episodes cataloging what robot arms can't do — grapes, feathers, deflated balloons — but today I want to talk about something that might actually be harder than all of them combined. Luna: You're going to say folding a fitted sheet, aren't you? Lucas: I am absolutely going to say folding a fitted sheet. And if you've ever stood in your bedroom wrestling with one, you know exactly why this is the Everest of robotic manipulation. Luna: I hate folding fitted sheets. I usually just ball it up and shove it in the linen closet. Lucas: That's what 90 percent of humans do. But even that balling-up motion — the crumpling, the compression, the irregular force distribution — is incredibly hard for a robot to replicate reliably. Luna: Okay, but we've had robots folding square towels for years. What makes a fitted sheet different? Lucas: Great question. The short answer is elasticity and asymmetry. A square towel is a flat, rectangular piece of fabric with straight edges and right angles. A fitted sheet has curved corners with elastic seams, a top and bottom that are different, and — critically — there's no single 'correct' fold. The geometry changes depending on how you grab it. Luna: So there's no consistent reference frame for the robot to plan around. Lucas: Exactly. And that's the core challenge. In 2025, a team at MIT CSAIL published a benchmark called — I'm not joking — 'Fitted Sheet Folding.' They tested four state of the art dual-arm manipulation systems on a set of thirty fitted sheets of varying sizes and elastic tensions. The best-performing system succeeded only 38 percent of the time. Luna: Thirty-eight percent? That's worse than my first attempt. Lucas: It's worse than a random flailing motion. The researchers found that the failure modes were all sensor-related: the tactile sensors couldn't distinguish between the elastic seam and the fabric body, the vision system lost track of the corners once the sheet bunched up, and the force-torque sensors couldn't tell if the sheet was being pulled taut or allowed to sag. Luna: I want to dig into those sensor failures, but first — why does it matter? I mean, how many fitted sheets are we talking about in the real world? Lucas: Enormous numbers. The global textile automation market — which includes industrial laundry, hospitality, and healthcare — was valued at roughly forty-seven billion dollars in 2025. And a big chunk of that is in folding and sorting. Hotels, hospitals, cruise ships — they all have to process thousands of fitted sheets every day. Right now, it's almost entirely manual labor. Luna: So there's a real economic incentive to solve this. It's not just a lab curiosity. Lucas: Not at all. A single large hotel chain might fold ten thousand fitted sheets per week across their properties. If a robot could do that even with ninety percent reliability, you'd save millions in labor costs. But we're nowhere near that. Luna: Walk me through the specific failure modes. What happens when the robot tries to pick up a fitted sheet? Lucas: First, the initial grasp. Most systems use a top-down suction gripper or a parallel-jaw grip. The sheet is usually presented as a pile on a table. The robot has to segment the pile into layers, pick up one sheet, and then locate the corners. But because the sheet is elastic and has seams, the suction often picks up multiple layers, or the gripper slips when the elastic contracts. Luna: So the very first step — picking it up — is already unreliable. Lucas: Right. And if you manage to pick up a single sheet, the next step is 'corner detection.' The robot has to identify the four corners of the fitted sheet. But the corners are curved, and the elastic makes the fabric pucker in unpredictable ways. The vision system might think a fold is a corner, or it might miss a corner entirely because it's hidden under another layer. Luna: And then the actual folding motion — that's a whole other problem. Lucas: Exactly. Once you have two corners in each gripper, you have to bring them together. But the sheet's elastic tension changes as you move it. The force profile is nonlinear. If the robot applies too much force, it stretches the sheet and loses alignment. Too little force, and the sheet droops and tangles. The MIT team found that even with perfect corner detection, the folding trajectory failed thirty percent of the time because of tension variation alone. Luna: So it's a simultaneous perception and control problem. You need to sense the elastic state in real-time and adjust your motion accordingly. Lucas: That's the research frontier. There's a group at Stanford working on what they call 'elastic-aware manipulation' — they embed tiny magnetic markers in the fabric and use a magnetic field sensor array to track deformation in real time. It works in the lab, but it's not scalable because you'd have to modify every sheet. Luna: Are there any non-vision approaches that might work better? Like purely tactile? Lucas: There's promising work from the University of Tokyo on 'active tactile exploration' — where the robot uses its fingers to feel for the elastic seam and then follows it to the corner. That bypasses vision entirely. But it's slow: it takes about two minutes per sheet, whereas a human can do it in about fifteen seconds. Luna: So we're trading speed for reliability. Not ideal for a hotel laundry operation that needs throughput. Lucas: Exactly. And that's the tension in the whole field. The robotics community is really good at solving narrow problems in controlled environments. But a fitted sheet is the opposite: it's a mass-produced object with intentional asymmetry, and it's designed to be handled by human hands with years of implicit experience. Luna: It's humbling, isn't it? A task that most of us consider annoying but trivial — and it's completely beyond the reach of our most advanced robots. Lucas: That's why I find this topic so fascinating. It reveals the gap between what we think robots can do and what they actually can do. And it's a reminder that some of the hardest problems in robotics are the ones we don't even think of as problems until we try to automate them. Luna: You know, talking about these real-world limitations — it actually connects to something I've been thinking about. We put so much work into making these episodes clear and ad-free, and listener support is what keeps it that way. Lucas: Yeah, it's true. If you get something out of these deep dives into what robots can't do — or the occasional thing they can — and you want to help us keep digging into these questions, listeners can support the show at buy me a coffee dot com slash fexingo. Luna: And that support really does make a difference. It lets us spend the time on episodes like this, where we can go deep into a single problem instead of just skimming the surface. Lucas: Agreed. So back to the fitted sheet — one more angle I want to explore. There's been some interesting work on using reinforcement learning to train robots to fold deformable objects in simulation, then transferring that policy to the real world. It's called sim to real transfer. Luna: I've heard about that for rigid objects like boxes. Does it work for fabric? Lucas: It's much harder. Simulating fabric with realistic elasticity and friction requires enormous computational resources. Even then, the simulation is never perfect. The researchers at NVIDIA's robotics lab trained a policy in simulation that worked well in simulation, but when they transferred it to a physical robot, the success rate dropped from eighty-five percent to twenty-two percent. Luna: That's a huge reality gap. What was the main cause? Lucas: Mostly the tactile sensing. In simulation, the robot has perfect knowledge of the fabric's state. In the real world, the sensors are noisy, the fabric slides differently, and the elastic properties vary with humidity and temperature. The policy hadn't learned to handle that uncertainty. Luna: So the solution might be to train in simulation with randomized parameters — domain randomization, right? Lucas: Exactly. And that's what the latest research is doing. They vary the fabric stiffness, the friction coefficient, the sensor noise — all within realistic ranges. And it does improve transfer. But even with aggressive randomization, the best reported real-world success rate for folding a fitted sheet is still under fifty percent. Luna: That's remarkable. We're in 2026, and we can land rockets on drone ships, but we can't fold a bedsheet. Lucas: It's a great reminder that progress in robotics is deeply uneven. Some problems that seem simple to humans turn out to be incredibly complex, while things we find hard — like playing chess or solving calculus — are trivial for machines. Luna: So what's the next step? Do you think we'll see a commercial fitted sheet folding robot in the next five years? Lucas: I think we'll see a purpose-built machine that doesn't try to mimic human folding. Instead, it might use a different process entirely — like a conveyor system that stretches the sheet flat, then uses a series of rollers and air jets to fold it. That's more like an industrial folding machine than a humanoid robot. Luna: So the lesson is: don't try to imitate human manipulation. Exploit the robot's strengths. Lucas: Exactly. And that's the theme that runs through so many of these episodes. The hardest robotic problems are often the ones where we insist on mimicking human strategies, when a completely different approach might work better. Luna: Well, I'm going to go home and try folding my fitted sheet with a little more appreciation for how hard it is. Lucas: And if you figure out a new method, let us know. We might do a follow-up episode.