Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Handle a Towel
Transcript
- Lucas: Luna, I want to start with something that sounds trivial but isn't. Go to your bathroom, grab a towel, and toss it in the air. Watch it land. Now try to pick it up with one hand, from one corner, without smoothing it out first. Luna: I mean, I can do it, but I can already tell you why a robot can't. The towel doesn't hold its shape. Lucas: Exactly. And that's the core of today's topic. After a hundred and ten episodes on everything from catching balls to folding fitted sheets, we haven't really tackled the humble bath towel. It's a perfect example of a class of objects called compliant or deformable — things that change shape when you touch them. Luna: And before we dive into the tech, I want to mention something. If conversations like this are useful to you, if you find yourself thinking about robot towels more than you expected, listener support is what keeps this show ad-free. You can help at buy me a coffee dot com slash fexingo. Lucas: It does make a difference. And we're grateful. So, a towel. Why is it so hard? Let's start with the mechanics. Lucas: A rigid object — say, a steel block — has a fixed geometry. A robot arm can plan a grasp using CAD models or point clouds. It knows exactly where the surfaces are. A towel has near-infinite degrees of freedom. The moment you touch it, the fabric bends, folds, and drapes in ways that depend on the exact force you apply. Luna: So the robot doesn't know where the towel is, in a sense. It sees a blob of pixels that changes every frame. Lucas: Right. And the gripper is a huge problem. Most industrial grippers are parallel-jaw — two flat plates that close on an object. For a rigid box that works fine. For a towel, the jaws just push the fabric aside or pinch a tiny fold that then slips out. Luna: There's a great paper from MIT CSAIL last year — I want to say it was presented at ICRA 2025. They tackled precisely this problem: picking up a towel from a pile using only a single overhead camera and a parallel-jaw gripper. Lucas: I remember that one. Lead author was Yufei Wang. Their approach was clever: instead of trying to model the towel's physics in real time, they used a simulator to generate thousands of examples of successful grasps on simulated towels, then trained a neural network to predict a good grasp point from a single image. Luna: And it worked — something like an 85 percent success rate on real towels, which was a big jump from prior work. But here's the thing: they only picked up the towel. They didn't fold it. Lucas: Folding is a whole other level. Because now you need to execute a sequence of manipulations that depend on the towel's current state, which you can only partially observe. Humans do this by leveraging prior knowledge about how fabric behaves — we know that if you lift one corner, the rest will hang down. A robot has to learn that. Luna: There's another approach that uses two grippers — essentially two robot arms — to mimic human hands. That's the setup at the lab in UC Berkeley, where they've been working on folding laundry since about 2018. Lucas: The Berkeley system, called the 'Towel Folding Robot' — not the most creative name — uses two arms with special grippers that have a kind of 'pinch and slide' motion. It can fold a towel from a neatly laid-out state in about two minutes. But it requires the towel to be initially spread flat on a table, which is a huge assumption. Luna: So the real challenge is the starting condition. A robot that can fold a towel from a crumpled pile is still an open research problem. Lucas: And that's where simulation is making a big push. NVIDIA's Isaac Gym, for instance, lets researchers train policies in simulation where they can run thousands of trials in parallel. The problem is that simulated fabrics are still approximations. Real fabric has friction, stiffness, and air resistance that are hard to model perfectly. Luna: There's a concept called the 'sim to real gap.' A policy that works perfectly in simulation often fails on the real robot because the towel in simulation is not quite the same as the towel in reality. Lucas: Researchers are tackling that with domain randomization — essentially, they vary the physics parameters in simulation so the policy learns to be robust to different fabric behaviours. It's the same technique that made OpenAI's Dactyl hand work for cube manipulation. Luna: But even then, a towel is more complex than a cube. The number of possible configurations is astronomical. Lucas: Let me give you a number. A single 2D polygon mesh of a towel with just 100 vertices has, roughly, 100 factorial possible configurations if you consider vertex positions. Obviously that's a gross overestimate, but it gives you an idea. The state space is enormous. Luna: So what's the pathway to a commercial laundry-folding robot? We've seen some startups — Foldimate, Laundroid — but they've all struggled. Lucas: Foldimate is actually still around, but they pivoted to a sort of semi-automated ironing board. Laundroid, the Japanese company, shut down in 2019 after selling only a few hundred units. The problem wasn't just the manipulation — it was reliability. The robots would jam, tear clothes, or mis-fold. Luna: And the cost was prohibitive. Laundroid was priced at around $16,000. For that, you could pay someone to fold your laundry for years. Lucas: Exactly. So the economics don't work yet. But the research is valuable beyond laundry. Towels, sheets, and clothing are a benchmark for a whole class of deformable objects — something that crops up in manufacturing, healthcare, and even space operations. Luna: In space, for example, you might need a robot to handle solar sails, which are essentially giant sheets of thin film. That's a compliant object. If you can fold a towel, you can deploy a solar sail. Lucas: That's a great connection. There's also the medical angle — handling surgical drapes, bandages, or even soft tissue. A robot that can manipulate a towel safely and precisely has implications for surgical robotics. Luna: So the towel problem is really the 'fruit fly' of manipulation. It's simple enough to study, but complex enough to reveal fundamental challenges. Lucas: I like that analogy. And like the fruit fly, progress in towel manipulation tends to lead to breakthroughs in other areas. The MIT paper I mentioned used a technique called 'GraspNet,' which is now being applied to other deformable objects like cables and bags. Luna: Cables are another nightmare. Same issue — infinite degrees of freedom, and they tend to tangle. Lucas: We could do a whole episode on cables. But today, let's stay on towels. One more thing I find fascinating: humans have an intuitive physics engine that lets us predict how a towel will behave. We don't think about it, but we're constantly doing mental simulation. Robots lack that. Luna: There's a concept called 'intuitive physics' in AI. Some researchers are trying to build neural networks that learn a physics model from observation, like a baby learns about the world. Lucas: And that's where deep learning and simulation come together. You can train a neural network to predict the next state of a towel given an action, using data from simulation. Then you use that predictor to plan a sequence of grasps and folds. Luna: But the predictor is only as good as the simulation. And simulation is still not perfect. Lucas: Right. So there's a loop: better simulation enables better policies, and better policies reveal where simulation is wrong, which drives improvements. Luna: I think we're going to see a commercial robot that can reliably fold a towel from a pile within the next five years. Not cheap, but possible. Lucas: I'd be cautiously optimistic. The hardware is getting cheaper — force-torque sensors and tactile skins are becoming more accessible. And the algorithms are improving fast. Luna: So what's the one thing you want listeners to take away from this episode? Lucas: That the towel is a perfect example of why robotics is still hard. We often think 'grasping' is solved, but it's only solved for rigid, predictable objects. The real world is messy, deformable, and uncertain. And towels, in their own quiet way, are a window into that frontier. Luna: Alright, next time you pick up a towel, maybe give it a little respect. Lucas: I know I will. Thanks for listening, and we'll see you next episode.