Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Wet Towel
Transcript
- Lucas: You know that moment when you grab a wet towel from the shower and you try to wring it out, and the water just — it doesn't all come out at once, it trickles and pools in one corner, and the fabric bunches unevenly? Luna: Yeah, and you end up with one soaking corner and the rest just damp. Lucas: Right. Now imagine teaching a robot arm to do that same wringing motion — but the robot has no concept of how water moves through fabric, no sense of weight distribution shifting in real time, and its grippers are basically two hard plastic fingers. Luna: I'm already picturing it dropping the towel or just squeezing air. Lucas: Exactly what happens. And that's the subject of a really interesting study that came out of MIT CSAIL in late 2025 — they tested twelve different industrial robot arms on the task of wringing and then folding a standard bath towel that had been soaked with a controlled amount of water, about forty percent of the towel's dry weight in water. Luna: Forty percent — so a towel that weighs, say, two hundred grams dry becomes two hundred eighty grams wet. That's a nontrivial weight shift. Lucas: It is. And the results were brutal — not one of the twelve arms could successfully wring the towel in under five attempts. The best performer, a Fanuc CRX-10iA with a two-finger parallel gripper, managed to wring out about sixty percent of the water on its fourth try, but it took over a minute and the towel ended up twisted into a knot that then it couldn't unfold. Luna: So the wringing itself sort of worked, but it created a new problem — the towel was now tangled. Lucas: Right. And that's the core challenge: wet deformable objects are a double problem. First, the object's shape changes continuously as water redistributes, and second, the gripper's grip force has to adapt because the towel gets heavier in some spots and lighter in others. Standard force-torque sensors on robot wrists update at about a thousand hertz, but that's still not fast enough to track the micro-shifts in water pooling. Luna: And vision systems can't see the water distribution either, right? Because water is transparent inside a colored fabric. Lucas: Exactly. RGB cameras see a wet towel as just a darker version of the same towel. Depth cameras get confused by the specular reflections off the water. So the robot is essentially blind to the very thing that's making the task hard. Luna: Which is why, I imagine, the researchers turned to simulation. I read a bit about this — they built a physics simulator that models fluid absorption into fabric fibers at a mesoscopic level, not just cloth as a mesh but actually water particles interacting with threads. Lucas: That's exactly what they did. The MIT team used a simulator called PlasticineLab, which is an open-source differentiable physics engine, and they added a fluid layer on top. They simulated a square of terry cloth with about ten thousand thread crossings and modeled water as a set of particles that could cling to those threads via surface tension parameters. Luna: And then they trained a policy in simulation and transferred it to the real robot? Lucas: They tried. And here's where it gets interesting — the sim to real transfer failed. The policy that worked perfectly in simulation, where water particles behaved ideally, collapsed in the real world within three seconds. The real towel's water didn't redistribute the same way because real fabric has irregularities — loose threads, uneven weave, different capillary action at the edges. Luna: So simulation is still too clean. It doesn't capture the messiness of a real towel. Lucas: Right. And that's a huge open problem. But there's a group at Carnegie Mellon that took a different approach — they didn't simulate the water at all. Instead, they used a method called 'implicit representation via tactile feedback.' They built a gripper with sixteen individual pressure sensors on each finger, and they trained a neural network to predict the water distribution purely from the pressure pattern as the gripper squeezed the towel in different places. Luna: So the robot learns to 'feel' where the water is by squeezing, not by seeing. Lucas: Exactly. And it worked — they got a robot to wring a towel with about seventy-five percent water extraction on the first attempt, which is better than any of the mit tested arms. But it took about four minutes of exploratory squeezing before the robot attempted the actual wringing motion. So the trade-off is speed. Luna: Four minutes to wring a towel — you could do it by hand in, what, ten seconds? Lucas: Maybe fifteen if you're being thorough. So it's not practical yet. But the principle is important: the robot is learning a model of the object's internal state through touch, and that's a skill that could generalize to other wet deformable objects — a wet sponge, a soaked mop head, a waterlogged cloth in a factory setting. Luna: And the industrial applications are real. Think about textile manufacturing, where fabric is dyed and then needs to be wrung and dried. Or cleaning robots in hospitals that need to handle soiled linens. Or even food processing — handling wet lettuce or soaked grains. Lucas: The lettuce example is a good one. There's actually a company in the Netherlands, ISO Group, that builds robotic systems for commercial kitchens, and they told me that handling wet lettuce after washing is one of their top customer complaints. The leaves stick together, they tear easily, and the weight changes so fast that the robot's grip force is almost always wrong. Luna: So the same physics problem shows up in a completely different context. Lucas: Yeah. And it's not just about water — any liquid absorbed into a deformable material creates this challenge. Oil in a shop rag, cleaning solution in a microfiber cloth, blood in a surgical sponge. Each has different viscosity, different absorption rates, different surface tension. Luna: And each would require retraining the tactile model, I assume. Lucas: In principle, yes, because the pressure patterns change. But the CMU group found that if you train on water and then fine-tune on a new liquid with just ten exploratory squeezes, the model adapts to about eighty percent accuracy. That's promising for generalization. Luna: So we're not at the point where a robot can walk into a laundry room and fold a basket of wet towels, but we have a path — tactile sensing combined with fast-adapting neural models. Lucas: That's exactly the state of play in mid-2026. The hardware is good enough — those Fanuc and KUKA arms have the dexterity and the sensors. The bottleneck is the software model of the wet object's internal state. And that's where I think the next breakthrough will come, probably from combining the tactile approach with a very lightweight simulation that runs in real time on the robot's onboard computer. Luna: A hybrid: real tactile data feeding a simplified physics model. Lucas: Exactly. There's already work on that from a team at UC Berkeley — they call it 'differentiable tactile rendering.' They basically treat the tactile sensor readings as a low-resolution image and then use a neural network to predict the underlying fluid distribution. It's not real-time yet, but it's close. Luna: And if that works, we might finally see a robot that can handle a wet towel without making a mess. Lucas: Which would be a big step. And you know, it's interesting — the towel problem is a perfect example of why robotics is so hard. It seems trivial to us, but it's a deep physics and perception challenge. And the fact that researchers are openly sharing their failures, like the MIT sim to real collapse, is actually really healthy for the field. Luna: Absolutely. And if any of this tech conversation gave you something useful — a new way to think about robot manipulation, or just a cool fact about wet towels — a couple of dollars a month is genuinely what keeps these shows going. Buy me a coffee dot com slash fexingo, if you've gotten something out of them. Lucas: Yeah, it really does make a difference. So — back to that Berkeley work. They're using a differentiable renderer that can backpropagate through the tactile sensing pipeline, which means the robot can learn to adjust its grip based on the predicted water distribution without needing explicit labels. Luna: So it's self-supervised learning applied to tactile data. Lucas: Exactly. And the early results show that after about a hundred real-world squeezes on a wet towel, the robot can predict the water distribution to within about fifteen percent accuracy — enough to choose a better grip point for wringing. It's not perfect, but it's a lot faster than the CMU approach of four minutes of exploratory squeezing. Luna: And how does it handle a different type of fabric? Like a microfiber cloth versus terry cloth? Lucas: They tested that — they trained on terry cloth and then tested on microfiber without any additional training. The accuracy dropped to about forty percent. But with just five additional squeezes on microfiber, it climbed back to seventy percent. So the model learns something about the general physics of wet fabric, not just the specific weave. Luna: That's encouraging. Because in a real factory or hospital, you'd have a mix of fabric types. Lucas: Right. And the researchers are now working on a version that doesn't even need those calibration squeezes — they want the robot to infer the fabric type from the first touch. They think they can do it by analyzing the initial pressure transient when the gripper first contacts the fabric, because different fabrics have different stiffness and compression profiles. Luna: So the robot learns to 'feel' the fabric type in the first millisecond of contact. Lucas: That's the idea. And if that works, you could imagine a robot that picks up a towel, instantly knows it's terry cloth, knows it's wet, knows roughly where the water is, and then executes a wringing motion that's optimized for that specific state. That would be a genuine breakthrough. Luna: And it would finally solve the towel problem — at least for one type of towel. Lucas: Yes. And from there, the methods would generalize to other wet deformable objects. The same tactile-first approach could work for wet sponges, wet rags, even wet food items like tofu or soaked bread — though those have their own challenges with fragility. Luna: Tofu is a whole other nightmare. But that's probably a future episode. Lucas: Definitely. For now, the takeaway is that the wet towel problem is a perfect microcosm of the broader challenge in robotic manipulation: the object's internal state is invisible to standard sensors, and the physics is too complex to simulate perfectly. The field is moving toward solutions that rely on rich tactile feedback and fast-adapting models, and that's a trend we'll see more of. Luna: And maybe in a few years, we'll all have a robot that can fold our laundry after a shower. Lucas: I'd settle for one that can wring out a mop head without leaving a puddle. But we're getting there.