Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Fold a Towel
Transcript
- Lucas: So, you can buy a robot arm today that'll pick a transmission off a conveyor belt, rotate it 180 degrees, and drop it exactly into a jig — cycle time under ten seconds, repeatable to within a tenth of a millimeter. Luna: Right. That's standard industrial robotics. What's the catch? Lucas: The catch is, hand that same arm a bath towel and ask it to fold the towel. It will fail. It might even fail spectacularly — dropping it, tangling it, or just freezing because its internal model of the world broke. Luna: And this isn't a niche problem. Laundry is the chore people say they hate most, after cleaning the bathroom. There's a huge market. Lucas: Exactly. So why is folding a towel harder than assembling a transmission? The short answer: a towel has effectively infinite degrees of freedom. A rigid object — a metal part, a box, a phone — has six degrees of freedom in space. You can model its position and orientation precisely. A towel can bend, wrinkle, stretch, fold in a thousand ways. Its state changes the moment you touch it. Luna: So the robot's brain — the planner — can't predict what the towel will look like after one grasp, because the grasp itself changes the shape. Lucas: That's the core difficulty. In robotics, we call this 'deformable object manipulation.' And it's been a research problem for decades. But recently, there's been real progress. A team at UC Berkeley's AUTOLab published a system called SpeedFolding in 2021. They used a dual-arm robot and a neural network trained purely in simulation — they never showed it a real towel during training. And it could fold a towel from a random crumpled state in under two minutes. Luna: Two minutes is slow compared to a human — I can fold a towel in maybe ten seconds — but the fact that it worked at all was a breakthrough. Lucas: Exactly. And the key innovation was in how they represented the towel. Instead of trying to track every pixel, they used a 'particle' representation — treating the towel as a cloud of maybe a hundred points, each with a position and a velocity. The neural network learned to predict how those particles would move when the grippers pulled or pushed. Luna: So it's not trying to see every fold, just a sparse model of the overall shape. Lucas: Right. That's the direction a lot of labs are going. But transferring from simulation to reality is still fragile. The SpeedFolding system worked reliably on towels of a certain size and fabric, but change the material — say, a silk scarf or a heavy bathrobe — and the failure rate spikes. Luna: And what about the grippers? I imagine standard parallel-jaw grippers — the pincher kind — are terrible for fabric. Lucas: Terrible is a generous word. They puncture, they slip, they can't sense whether they've grabbed one layer or four. So researchers have developed custom grippers. One approach uses adhesive pads — like a weak glue — to lift a single layer. Another uses electrostatic attraction, basically a charged surface that grabs fabric without mechanical force. There's even a gripper from MIT CSAIL that uses tiny needles — like a bed of microscopic hooks — to engage the fibres without damaging them. Luna: I've seen that. It's called a 'microneedle gripper.' It looks like a brush, but the bristles are angled so they catch the fabric when pulled. Luna: There have been commercial attempts, right? I remember a startup called FoldiMate, and another one called Laundroid. Whatever happened to them? Lucas: Laundroid went bankrupt in 2019. FoldiMate pivoted and eventually shut down. The problem was twofold. First, the hardware was too expensive — these machines cost thousands, and consumers weren't willing to pay that for a single-task robot. Second, the reliability was never good enough. If a robot fails to fold one shirt out of ten, you still have to finish the job manually — so what's the point? Luna: But the industrial side is different, I'd imagine. Hotels, hospitals, laundromats — they have volume and consistency. If a robot could fold a thousand sheets a day, even with a 5 percent failure rate, you'd still save labour costs. Lucas: That's where the real market is, and a few companies are going after it. One is a Japanese firm called Seven Dreamers — they're not the same as the old Laundroid — they focus on commercial laundry. Another is a us based startup called Monotask, which is building a sheet-folding robot for hotels. Their approach is to simplify the problem: they only handle flat sheets, not fitted ones, and they use a custom conveyor system that aligns the fabric before the robot touches it. Luna: So they reduce the degrees of freedom by constraining the environment. That's clever — you don't need general intelligence if you can control the inputs. Lucas: Exactly. And that's the pattern across deformable object manipulation. The most successful systems are the ones that narrow the problem. For example, in apparel manufacturing — sewing garments — robots are starting to handle fabric, but only after the fabric is precut and laid flat on a table with vacuum hold-down. They don't pick up a random pile; they pick from a known pose. Luna: So the holy grail — a robot that can take a wrinkled heap of laundry from a basket and fold it perfectly — is still a research problem, not a product. Lucas: It is. And it might stay that way for a while. But the research is accelerating. In the past two years, there's been a wave of papers using diffusion models — the same kind of AI that generates images — to plan folding motions. Instead of predicting one optimal path, the model generates a distribution of possible motions, then picks the most likely to succeed. Luna: Interesting. So it's probabilistic, not deterministic. Lucas: Right. And that's important because with deformable objects, there's never a single right answer. The robot has to adapt in real time. There's also work on tactile sensing — equipping grippers with pressure sensors so they can feel whether they have a single layer or a fold. That closes the loop between perception and action. Luna: Feeling instead of seeing. That makes sense for opaque fabric. Lucas: And for dark fabric, or fabric with patterns that confuse cameras. So the field is moving. But I think the lesson for now is: folding a towel is a benchmark for robot dexterity. If we can solve it robustly, we'll have solved a whole class of manipulation problems — handling cables, medical gauze, food items like lettuce leaves. Each of those is a multi billion dollar industry. Luna: And yet my robot vacuum still can't handle a shoelace on the floor. It just eats it. Lucas: That's a different problem — but related. Actually, the shoelace problem is a great example of deformable object avoidance, not manipulation. The robot sees a long thin thing and doesn't know if it's a cable or a worm, so it tries to go over it and gets tangled. Luna: Right. So perception and planning are still the bottlenecks. But for the folding problem, what's the single biggest hurdle left? Lucas: I'd say it's the 'first grasp' problem. If the robot has to pick a towel from a random pile, it has to decide where to grab. Grab the wrong spot — the middle of a fold — and you'll end up with a twisted mess. There's research at Carnegie Mellon using 'next best grasp' algorithms that look at the current configuration and predict where a second grasp would be most effective. But it's still slow. And speed matters because in a commercial setting, you need to fold one towel every few seconds, not every few minutes. Luna: So we're probably years away from a reliable laundry robot for the home. But maybe five years for hotels. Lucas: That's my guess. And it's worth noting that the hardware cost is dropping. Cheap, force-sensitive arms like the Franka Emika Panda or the Universal Robots e-Series make experimentation accessible. The research community is collaborating more. There are now benchmarks like the 'Folding Challenge' at ICRA — the big robotics conference — where teams compete to fold a standard towel in under three minutes with no human intervention. Luna: What's the current record? Lucas: As of last year, the best system could fold a towel in about 90 seconds with a success rate of 85 percent. That's impressive, but for a hotel, you'd need 99.9 percent reliability and a cycle time under 20 seconds. So there's still a gap. Luna: But the progress from 2015 to now is huge. Back then, no robot could reliably fold any towel from a random configuration. Lucas: You're right. And the same trajectory is happening in other areas of deformable manipulation. Cable routing for wiring harnesses — that's a massive problem in automotive and aerospace. Boeing has a research group dedicated to robot cable manipulation. They use a similar particle-based representation to guide a robot arm that routes wires through a fuselage. It's still slower than a human, but it's consistent, and it doesn't get carpal tunnel. Luna: And that's a $100 billion industry — aerospace wiring. So even incremental progress there pays off. Lucas: Huge payoff. And the same techniques apply to surgical robotics — tying sutures, manipulating tissue. Deformable object manipulation is the next frontier after rigid object manipulation. We've nearly solved rigid. Now the hard part begins. Luna: Before we wrap, I want to circle back to something you said earlier about the research community. It strikes me that a lot of this work is open-source — the code, the datasets, even the gripper designs. That's a pretty big deal for accelerating progress. Lucas: It is. And that's actually a good moment to mention something about how this show works. We keep the Robotics Podcast ad-free — no sponsors, no mid-roll interruptions — and that's only possible because a small group of listeners contributes monthly through buy me a coffee dot com slash fexingo. Luna: Yeah, and it's genuinely what funds the research we do to bring you these deep dives. Every episode takes hours of reading papers and talking to engineers. Lucas: So if you've found today's conversation useful and you want to see more like it, that's the link. No pressure, no perks — just keeping the podcast independent. And if you can't, that's fine too. The episodes will keep coming either way. Luna: Absolutely. Now, back to folding. You mentioned surgical sutures — that's a whole other level of precision. How far along are those systems? Lucas: They're in early research stages. The da Vinci surgical robot can help a surgeon tie knots, but fully autonomous suturing is not there yet. The challenge is similar: tissue deforms unpredictably, and a needle puncture changes the local geometry. But there's promising work at Johns Hopkins using reinforcement learning to learn suturing motions in simulation, then transfer to a physical setup with a phantom tissue model. They achieved about 80 percent success on a simple running stitch. Luna: So we're seeing a pattern: simulation, particle representations, and tactile sensing are the three pillars. Lucas: That's exactly right. And I think in the next five years, we'll see the first commercial systems that handle one specific deformable object — hotel sheets, maybe, or surgical gauze — with sufficient reliability to be deployed. The general-purpose laundry-folding robot for the home is probably a decade away. But the building blocks are being laid now. Luna: And I guess that's the exciting part. Every towel folded in a lab today is a step toward the future we were promised decades ago. Lucas: Yeah. The robot that folds your laundry is coming. It just needs to learn how to grab that first corner.