Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Inside the Hand That Folds a Towel Without Crushing It
Transcript
- Lucas: So a few years ago, a team at MIT taught a robot arm to fold a towel. It took about a minute and a half per towel, and it used a two-fingered gripper that you could buy off the shelf. No custom end effector, no vacuum arrays. Just a wrist sensor and some clever planning. Luna: And that still sounds like a big deal, because folding a towel is one of those tasks that feels easy to us but is genuinely hard for a machine. The fabric moves, it stretches, it has no fixed shape. Lucas: Exactly. And what's wild is that even after all the advances in warehouse automation and bin picking, deformable objects like fabric remain the frontier. The challenge is that a towel isn't a rigid body. It has infinite degrees of freedom, and the robot can't just grab it once and assume it'll hold its shape. Luna: So what changed at MIT? Was it a vision breakthrough or something else? Lucas: It was really a sensing breakthrough. Instead of relying only on cameras, they added a force-torque sensor at the wrist. That lets the robot feel the tension in the fabric as it pulls. It's almost like the difference between watching someone fold a towel and feeling the fabric yourself. Luna: And feeling matters because you can't tell from a camera alone whether you've got a single layer or two layers bunched together. Lucas: Right. With force feedback, the robot can detect the resistance and adjust mid-fold. It's not just following a fixed trajectory. It's responding to the physical state of the towel in real time. Luna: That sounds like a different paradigm from the old pick and place. More like a human hand that feels its way through the task. Lucas: Precisely. And that's why I wanted to dig into this specific example today, because it illuminates the broader shift in robotics from rigid, pre-programmed motions to compliant, sensor-driven manipulation. Luna: And when you say 'compliant,' you mean the robot is willing to give a little, not just force its way through. Lucas: Exactly. Compliance is about allowing small deviations. If the fabric resists, the robot doesn't just plow ahead. It reads the force and adapts. That's the key to handling anything deformable, from a towel to a cable to a piece of raw chicken. Luna: And that's why we've done so many episodes on why robot arms fail at these tasks. The common thread is that they lack that tactile feedback. Lucas: Right. And the MIT towel-folding work is a concrete proof point that when you add the right sensing, those tasks become tractable. Now, the robot isn't as fast as a human yet, but it's reliable enough to be interesting commercially. Luna: So who's actually trying to commercialize this? Is anyone building a laundry-folding robot that we might see in our homes? Lucas: There are a few companies, but the MIT team spun out a startup called FoldiMate — wait, no, that's the one that had a booth at CES. The MIT spinout is actually called 'Foldit' — no, hold on. Let me get this right. The team behind that specific research, led by a researcher named Yahav Avigal, later started a company called 'Foldit' — no, I'm confusing it with the protein-folding game. The company is called 'Laundroid' — no, that's a different one. Let me just say there are several companies in this space, but the MIT research itself is a great illustration of the core principles. Luna: Fair enough. But you're saying the principles are the real takeaway, not any one product. Lucas: Exactly. Because the same force-feedback approach is being applied to other deformable objects. Think about wiring harnesses in automotive manufacturing, or fabric in the apparel industry. The moment you can feel what you're handling, a whole new class of tasks opens up. Luna: And that's a big deal because the global apparel market is huge, and automation there has lagged behind other sectors for decades. Lucas: Right. And it's not just about folding. It's about sewing, which is even harder. But the same sensing principle applies. Luna: So let's talk about the actual sensor. What does a force-torque sensor do, exactly, and why is it so key? Lucas: It measures the six components of force and torque: three linear and three rotational. In the MIT setup, it sits between the robot arm and the gripper, so every time the gripper touches something, the sensor reports the force vector. The robot can then infer things like 'I've grabbed the corner' or 'there's too much tension on this side.' Luna: And that's different from tactile skin sensors that measure pressure across a surface. Lucas: Right, and both have their place. The wrist sensor gives you an overall sense of the force, while tactile skin gives you local contact information. The MIT team used the wrist sensor for the folding task, but many advanced grippers combine both. Luna: And cameras still matter, right? They're not obsolete. Lucas: No, vision is still essential for initial perception, like finding the towel on a table. But vision alone can't tell you if you're pinching two layers or one. That's where force sensing shines. Luna: So it's a fusion of vision and touch. And that's where the field is heading. Lucas: Absolutely. And this brings me to something that might be useful for you as a listener. If today's tech conversation gave you something usable — maybe a mental model for why robots struggle with certain tasks — the way these shows stay ad-free is listener support. If you'd like to keep that going, you can buy me a coffee at buy me a coffee dot com slash fexingo. It's a small gesture, but it helps us keep digging into these topics without sponsorship noise. Luna: And I'd say it's worth it, especially since we get to geek out about sensor fusion and fabric physics without cutting to a commercial mid-explanation. Lucas: Exactly. So, back to the folding. One of the most interesting parts of the MIT approach is how they handle the planning. They don't just script a single fold sequence. They use a model that predicts the state of the towel after each manipulation, and the robot re-plans based on the actual measured state. Luna: So it's more like a control loop than a one-shot program. Lucas: Precisely. And that's the key to robustness. If the towel slips a bit, the robot can compensate. That's something we humans do naturally, but it's a big deal for robots. Luna: And that robustness is what makes it commercially viable. Because in a real factory or home, things won't be perfectly staged every time. Lucas: Right. And that's why the big automation companies are investing heavily in force-controlled robots. Universal Robots, for example, has made compliance a standard feature on their collaborative arms. It's not just a lab curiosity anymore. Luna: So if I'm a manufacturing manager, what's the practical takeaway from today's episode? Should I be looking at force-torque sensors for my line? Lucas: If you handle any kind of deformable material or need precise assembly, absolutely. The cost of these sensors has come down significantly over the past few years. You can get a decent one for a few thousand dollars, which is nothing compared to the cost of a robot cell that can't adapt to variation. Luna: And that's the same reason why the automotive industry has been early adopters of force-controlled assembly, right? Lucas: Exactly. Inserting a gearbox or a bearing into a tight fit requires delicate force control. If the robot pushes too hard, it stalls the motor; if it doesn't push enough, the part doesn't seat. Force-torque sensing solves that. Luna: So it's not just about towels. It's a general capability that's spreading across industries. Lucas: And the towel-folding demo was a great proof of concept because it's so intuitive to understand. Everyone knows folding a towel is easy for us but hard for machines. When a robot finally does it, you know something fundamental has changed. Luna: And there's still plenty of room to improve. The MIT robot took a minute and a half, and a human takes maybe thirty seconds. So the gap is shrinking, but it's not closed. Lucas: Right. And the next frontier is even more complex deformable objects, like cloth that's wrinkled or has variable thickness. But the research is moving fast, and I wouldn't be surprised if in a few years we see laundry-folding robots in commercial laundromats. Luna: That's a future I'm curious to see. In the meantime, I think I'll stick to folding my own towels, but I'll appreciate the complexity a bit more. Lucas: And maybe you'll look at your robot vacuum a little more forgivingly next time it tangles in a curtain. Luna: Ha, fair point.