Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Fold a Fitted Sheet
Transcript
- Lucas: There's a YouTube rabbit hole I fell into last weekend. A guy spends three minutes trying to fold a fitted sheet perfectly — tucking corners, smoothing seams, the whole ritual. And he finally nails it. Then he holds up the sheet, triumphant, and you see it's still basically a lumpy rectangle. That's when I thought: a fitted sheet is the ultimate test for a robot arm. Luna: Because it's not just fabric — it's elastic fabric with pockets. Every corner is a different orientation, and the geometry changes the moment you touch it. Lucas: Exactly. And that's what makes it so hard for robotics. Most industrial robot arms thrive on rigid, predictable objects — a metal bracket, a plastic housing, a box that's always the same shape. A fitted sheet is none of those things. It's what the field calls a 'highly deformable object,' and it breaks almost every assumption in classical manipulation. Luna: So where does the research stand right now? I know Berkeley's AUTOLab has done work on cloth folding — they had a paper maybe three years ago on folding towels with a two-arm setup. Lucas: Right — that was the 'learning to fold' paper from 2016, actually. They used a bimanual robot and a physics simulator to teach the system how to grasp and fold a towel into a rectangle. It worked, but the towel was a simple square of cotton — no elastic, no fitted corners. Since then, teams at MIT's CSAIL and Carnegie Mellon have pushed into 'deformable object manipulation' with more complex cloths. But a fitted sheet is the nightmare scenario. Luna: Let's break down why. First problem: perception. How does the robot even see what it's holding? Lucas: Great place to start. Standard depth cameras — like an Intel RealSense or an Azure Kinect — can capture point clouds of a surface. But a deformable sheet has no fixed shape, so the point cloud changes every frame. The robot can't rely on a pre-loaded CAD model. It has to infer the current state from vision alone. And with an elastic sheet, the camera might see a fold that disappears when the robot moves its arm. Researchers use something called 'active perception' — the robot pokes the cloth, watches how it deforms, and updates its model. But that adds time and uncertainty. Luna: Then you have the gripper problem. Most robot grippers are either parallel-jaw — two rigid fingers — or suction cups. Neither is great for a floppy elastic corner. Lucas: Right. A parallel-jaw gripper pinches a point, but the fabric around it still droops unpredictably. Suction cups work on smooth, non-porous surfaces — but a cotton sheet lets air through, so the seal fails. There's been work on 'soft grippers' — silicone fingers that conform to the shape — but they lack the precision to tuck a corner into an elastic pocket. A 2022 paper from MIT proposed a 'fabric-friendly gripper' using micro-spines, like tiny hooks that grab individual threads. It works on denim, but on a thin sheet, it can tear the material. Luna: And even if you could grip it, folding requires two hands. You need one arm to hold the sheet, another to perform the tuck. Bimanual manipulation is its own challenge — coordinating two arms without collision, especially when the object between them is constantly changing shape. Lucas: That's exactly the bottleneck. In the Berkeley towel-folding work, they used two arms and a 'trajectory library' — pre-computed motion paths for common folds. For a fitted sheet, the number of possible states is enormous. You'd need a library of hundreds of thousands of trajectories, or a real-time planner that can simulate cloth physics at 60 frames per second. Current physics engines — like MuJoCo or PyBullet — can simulate a rigid body in microseconds, but a cloth mesh with hundreds of vertices takes milliseconds per step. That's too slow for real-time control. Luna: So a human can fold a fitted sheet — badly — in about a minute, using intuition and muscle memory. A robot would need either perfect perception, perfect physics, or both. What's the state of the art? Lucas: The closest I've seen is work from CMU's Robotic Manipulation Lab on 'garment folding from a single demonstration.' They use a neural network trained on simulated cloth data to predict the fold configuration from a single image. Then they execute a learned policy to manipulate the garment. It works on t-shirts — folding them to a target shape with about 80 percent success. But t-shirts have a defined structure — sleeves, collar, seams. A fitted sheet is just a continuous elastic surface with four ambiguous corners. Luna: And those corners — the elastic bands — they're the killer. They add tension. When you pull one corner, the opposite corner pulls back. The robot has to model the tension field across the entire sheet to avoid bunching. Lucas: Exactly. A 2024 paper from Stanford's Dexterous Manipulation Lab addressed this with 'tension-aware grasping.' They instrumented a sheet with markers and used a force-torque sensor on the wrist to measure strain. The robot could then adjust its grip to minimize wrinkles. But it required the exact sheet to be pre-mapped with markers. In the real world, your robot doesn't know the elastic modulus of your specific sheet. Luna: So we're a long way from a robot that can fold laundry. But is there a commercial incentive? I mean, Amazon's warehouses still use humans for tasks like sorting soft goods. Lucas: That's actually a great point. Amazon has an entire research group — Amazon Robotics — working on 'deformable object handling.' They've deployed robots for rigid items, but for clothing, they still rely on human pickers. The reason is simple: the cost of a robot that can reliably fold a fitted sheet exceeds the labor cost. But if you look at the demographics — aging populations, labor shortages in care industries — there's a growing market for home robots that can do laundry. And laundry is the most hated chore, according to multiple surveys. Luna: Speaking of things that are worth it — if this conversation gave you something useful, honestly, if it was worth a coffee to you, that's the link: buy me a coffee dot com slash fexingo. Just the smallest gesture keeps the show ad-free and independent. Lucas: And that really does make a difference — we don't run ads, and listener support is the only reason we can spend time digging into these niche robotics questions. So thank you if that's something you do. Luna: Alright, back to the sheet. You mentioned that the tension problem is being tackled. Any promising approaches? Lucas: One direction I find really clever is 'memoization of folding trajectories.' Instead of planning from scratch each time, the robot builds a library of successful folds for different initial configurations. It uses a 'visual similarity metric' to match the current sheet state to the closest stored example, then replays the trajectory with small corrections. This was demonstrated on towels in 2021, and researchers are extending it to fitted sheets by adding an elastic model. The key insight is that while the sheet can bend infinitely, the number of 'useful' fold paths is actually finite — because you're always aiming for the same final rectangle. Luna: So the robot learns a set of muscle memories, like a human does. That's promising. But what about the perception of the elastic corners? They're the hardest to locate if the sheet is crumpled. Lucas: Right. A 2023 paper from MIT CSAIL introduced 'CornerNet for Cloth' — a convolutional neural network trained on synthetic images of crumpled sheets with labeled elastic corners. They generated 50,000 random crumple configurations in simulation. The network achieves 94 percent accuracy in identifying all four corners from a single overhead view. But then the robot has to reach those corners without dragging the sheet, which requires careful path planning. They used a 'reachability map' — basically a precomputed set of arm configurations that avoid collisions with the sheet itself. Luna: So the pieces are coming together — perception, simulation, control. But the integration is still fragile. Where do you see the first practical application? Lucas: I think it won't be a general-purpose laundry robot. It'll be a specialized machine — maybe in a hotel or hospital setting, where the sheets are uniform and the environment is controlled. A 2025 startup called FoldLogic is actually building a commercial fitted-sheet folder for hospitality. They use a custom frame that stretches the sheet taught, then a robot arm tucks and folds in a fixed sequence. It's not general — the sheet has to be placed in a specific orientation — but it works for 90 percent of standard king and queen sheets. That's the path: constrain the problem, then generalize. Luna: So the lesson is that a fitted sheet isn't just a hard problem — it's a litmus test for the entire field of deformable manipulation. If you can fold a fitted sheet, you can probably handle any fabric-based task in manufacturing or healthcare. Lucas: Exactly. And that's why I find it so fascinating. It's not just about clean laundry. The same algorithms could handle surgical gauze, carbon-fiber layups, tent assembly — any situation where flexible materials need precise manipulation. Every robotics lab that cracks this problem is opening a door to a whole new class of automation. And it all starts with that lumpy, elastic, impossible rectangle. Luna: I'm never going to look at my laundry pile the same way again.