Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Fold a Towel
Transcript
- Lucas: So, we've spent over a hundred episodes asking why robot arms can't do everyday things. And today, I want to tackle one that feels almost insulting in its simplicity. Luna: Let me guess. It's not folding a fitted sheet, because we did that one. Lucas: No, it's worse. It's folding a bath towel. You know, the rectangle you pull out of the dryer, give a little snap, and fold into thirds in about four seconds without thinking. Luna: Right, and a robot arm, with all its precision and force control, turns into a confused mess. Why is that? Lucas: Because a towel is a deformable object. Its shape isn't fixed. When a robot grabs a rigid object, like a metal bracket, it knows the exact geometry. Every point has a fixed position relative to every other point. A towel, though, bends, stretches, and drapes. Its shape changes the moment you touch it, the moment it moves, even just from gravity. Luna: So the robot's internal model of the world just falls apart. Lucas: Exactly. Most industrial robots rely on knowing the object's pose, its position and orientation in space, with millimeter accuracy. With a rigid part coming down a conveyor, you scan it, you know where it is, you pick it up. A towel has an infinite number of possible configurations. You can't just scan it and know where the corners are, because they're moving. Luna: And the corners are the key, right? Folding is basically matching the corners. Lucas: That's the core challenge. To fold a towel neatly, you need to identify the four corners, then bring two of them together precisely. But a wrinkle or a fold in the middle can make a corner appear offset by centimeters. And the robot's sensors, even high-res cameras and depth sensors, struggle to see through the folds. It's a perception problem as much as a manipulation problem. Luna: And if the robot's grip is even slightly off, the towel slips, and the whole thing is ruined. Lucas: Right. There's also the issue of friction. Towels are fluffy and absorbent, which is great for drying off, but terrible for robotic gripping. The fibers create a lot of unpredictable friction, and the fabric can slide between the gripper fingers in ways that are hard to model. Plus, the towel's weight distribution shifts as it folds, so the robot's force control has to constantly adjust. Luna: So it's not like the robot is just dumb. It's that the physics are genuinely complex. Lucas: And that complexity is why, for decades, most research focused on rigid objects. The math is clean, the models are tractable. Deformable objects were seen as too messy, too hard to generalize. But that's starting to change, partly because of the potential market. Luna: The market for laundry-folding robots. There's a real product out there, right? That one that looks like a giant drawer. Lucas: There is. One company makes a machine that uses a combination of air jets, belts, and mechanical arms to fold clothes. It's about the size of a large dresser, and it costs somewhere in the low four figures. It's not a traditional robot arm, but it tackles the same problem. And it's actually getting decent reviews, though it struggles with larger items like sheets and comforters. Luna: So there's progress, but it's not just about the hardware. What about the software? How are researchers actually teaching robots to fold towels? Lucas: A lot of it is coming down to simulation and reinforcement learning. Instead of hand-coding every possible towel configuration, researchers create virtual towels in a physics simulator. They give the robot a reward signal when it gets closer to a successful fold, and let it try millions of times, learning through trial and error. The challenge is making the simulation accurate enough that the robot can transfer what it learns to the real world. Luna: And that's the tricky part, the sim to real gap. Because the physics engine is never perfectly accurate. Lucas: Right. A simulated towel might behave slightly differently than a real one, and the robot's learned policy might not transfer. To bridge that gap, researchers use a technique called domain randomization. They randomize the simulation parameters, like friction, stiffness, and lighting, so the robot learns a more robust policy that works across a range of conditions. It's like making a student study from a textbook with slightly different versions of each problem. Luna: That's clever. But even with that, the robot still has to perceive the towel's state in real time. How do you track a deformable object that's constantly changing shape? Lucas: That's one of the active research areas. Some groups use dense depth sensors and fit a mesh model to the point cloud. Others use object simulators to track the towel's pose, essentially running a fast simulation alongside the real robot to predict where the fabric is. It's computationally heavy, but it's improving. And there's also work on special grippers, like ones with fingers that have tactile sensors, so the robot can feel the fabric between its fingers. Luna: I remember reading about a project that used two grippers to smooth out a towel by making small adjustments based on touch. It looked painfully slow, but it worked. Lucas: Slow and steady is the current state of the art. A robot might take a minute or two to fold a single towel, versus a human's four seconds. But the pace is accelerating. And honestly, if today's conversation gave you something to think about, we'd love to keep bringing you these deep dives. If it was worth a coffee to you, there's a link at buy me a coffee dot com slash fexingo. Every little bit helps keep this show going. Luna: Yeah, it's a small way to support the kinds of questions we dig into. And it's not like we're asking for a monthly subscription or anything. Lucas: Right, just a one-time cup of coffee if you feel like it. Okay, back to the towels. So there's a real economic incentive here. The home robotics market is projected to grow, and laundry is one of the most hated chores. If a robot could reliably fold a basket of mixed laundry, that could be a game-changer for busy households. Luna: But we're not there yet. And one of the biggest hurdles is dealing with mixed loads. A cotton t-shirt behaves very differently from a silk blouse or a wool sock. Lucas: Exactly. A robot trained on towels might not know what to do with a pair of jeans. So researchers are working on more generalizable policies, using things like meta-learning, where the robot learns how to learn new fabrics quickly. It's still early, but there are some promising results in handling novel objects with just a few demonstrations. Luna: So, in the near term, what's more likely: a dedicated towel-folding machine, or a general-purpose robot arm that can handle any fabric? Lucas: I think we'll see specialized machines first, like the one I mentioned, because they can be optimized for a narrow task. They don't need to be general. They just need to fold towels really well. And they can use cheaper, less precise components because the design constraints are tighter. A general-purpose arm has to be more flexible, which means more sensors, more computing, and a much higher price tag. Luna: And for most people, a four-thousand-dollar machine that only folds towels is a hard sell, unless they're really rich or really lazy. Lucas: Right. But as the tech matures, the cost could come down, and the functionality could expand. The same underlying algorithms for deformable object manipulation could also be used for tasks like handling cables, assembling clothing in factories, even surgical robotics, where manipulating soft tissue is critical. So the research has broader implications beyond just your laundry room. Luna: That's a good point. The medical angle is huge. Surgeons have to deal with deformable tissues all the time, and robotic surgery is already a thing. But those robots are still teleoperated by a human. Auto-mating some of those manipulations could be a big deal. Lucas: Definitely. But for now, the humble towel remains a benchmark. Every year, there are robotics competitions where labs try to fold a basket of laundry as fast and as neatly as possible. The times are getting better, but they're still not close to a teenager who just wants to get it done. Luna: So if I see a robot folding towels in my lifetime, what's the timeline? Lucas: I'd say within the next decade or so, we could see a reliable laundry-folding robot that costs less than a high-end smartphone. But it'll probably be a specialized appliance, not a general humanoid. The humanoid robots that are all over the news right now, they're still struggling with basic tasks like opening doors. Folding towels is way harder. Luna: That's a nice perspective. We've spent all this time on something that seems trivial, but it's actually a frontier in robotics. Lucas: And that's what makes it so fascinating. The things we do without thinking are often the hardest to replicate. They require not just precision, but an understanding of the physical world that we take for granted. So the next time you fold a towel, maybe give yourself a little credit. You're doing something that still stumps the most advanced robots on Earth.