Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Can't Wipe a Spill with a Cloth
Transcript
- Lucas: So we talk a lot on this show about things robot hands still can't do. And a lot of those tasks involve precision — threading a needle, folding a towel, handling a raw egg. But recently I came across a paper that made me realise there's a whole category of failure we've mostly ignored. Luna: What category is that? Lucas: Wiping. Specifically, wiping a liquid spill with a cloth. It sounds trivial. You grab a paper towel, you drag it across the counter, and the water goes away. But for a robot, that action involves at least three things that are shockingly hard to get right. Luna: Okay, I'm already curious. What are they? Lucas: First, you have to detect the spill and its boundaries — that's a vision and tactile problem. Second, you have to pick up a deformable cloth, which we've already covered in earlier episodes as a nightmare. But third — and this is the one I hadn't thought about — you have to manage the transition from dry cloth to wet cloth while maintaining just the right pressure and angle. Luna: Wait, why does the cloth getting wet change anything? I mean, the robot's still holding it. Lucas: That's exactly what the researchers found. Once the cloth starts absorbing liquid, its weight shifts, the friction against the surface changes, and the fabric itself becomes stiffer or more slippery depending on the material. A robot that was perfectly calibrated for a dry cloth suddenly loses its grip or applies too much force, so it either tears the wet paper towel or leaves streaks. Luna: So it's not just about sensing the spill — it's about adapting to a dynamic object that changes in your hand in real time. Lucas: Exactly. And there's a 2025 study from a joint team at MIT and the University of Tokyo that quantified this. They ran a standardised test where robots had to wipe a 10-centimetre puddle of water off a flat surface using a single paper towel. The success rate across six different grippers and three control algorithms: 32 percent. Humans in the same setup: 94 percent. Luna: That's a huge gap. So what's the specific failure mode? Is it the grip, the motion, or the absorption itself? Lucas: All three, but the biggest one was absorption timing. When a dry cloth first touches water, it wicks liquid rapidly into the fibres. That changes the center of mass of the cloth in the gripper. Most robotic controllers assume a constant load, so they compensate incorrectly. The robot either lifts the cloth too early, leaving half the puddle untouched, or it pushes too hard and just spreads the water around. Luna: I've seen Roomba mops do that — they just smear the dirt in a bigger circle. Lucas: Right — that's exactly the same problem at a lower complexity. But even advanced research arms struggle. The MIT team used a tactile sensor array that could measure local pressure at 120 points per square centimetre, and they still couldn't get the wipe to be consistent across different cloth types. Microfibre, paper towel, cotton rag — each one has a different absorption curve. Luna: So the robot has to learn the material properties of the cloth while it's already in contact with the liquid. That's an identification problem happening in milliseconds. Lucas: Exactly. And the motion itself isn't simple either. A human wipes by varying the angle of the wrist and the pressure as the cloth gets wetter. The first pass might be at 30 degrees with light pressure. As the cloth saturates, you increase the angle and push harder to wring it out against the surface. Robots don't have that intuitive control loop. Luna: It reminds me of the folding towels problem — but at least with folding, the cloth is dry and the geometry is predictable. With wiping, you've got fluid dynamics, absorption, and contact mechanics all at once. Lucas: Exactly. And I'll add one more layer: the spill itself isn't static. On a porous surface like wood, the water starts seeping in within seconds. So the robot has to wipe quickly but not so hard that it damages the surface. On a non-porous surface like stainless steel, the water beads up, so you need a different strategy — a sweeping motion rather than a pressing one. Luna: And all of this matters because... well, if we ever want robots to clean our homes or industrial floors, they have to handle spills. It's not a niche task. Lucas: It's actually one of the most common household cleaning actions. A 2024 survey by the Robotics Industries Association found that 'wiping up a spilled liquid' was the second-most desired feature in a home robot, after vacuuming. And yet no consumer robot today can do it reliably. Luna: So what's the most promising approach? Are researchers getting closer? Lucas: Yes — and it's interesting. The MIT team took a learning-based approach. They mounted a standard two-finger gripper on a Kuka arm and added a wrist-mounted force-torque sensor. Then they collected thousands of wipes with different cloths, different liquids, and different surfaces, and trained a neural network to predict the optimal pressure and angle based on the real-time force feedback. Their best model got the success rate up to 68 percent. Luna: That's better, but still not human-level. And that was in a controlled lab with known cloths and spills. Lucas: Right. When they introduced a new type of cloth — a synthetic chamois that the robot had never seen — the success rate dropped to 41 percent. So the model generalises poorly. The chamois has a much higher absorbency and a different friction coefficient, so the learned pressure profile was all wrong. Luna: So it's a classic overfitting problem. The robot learns the specifics of the training set, not the underlying physics of absorption. Lucas: That's the challenge. Some groups are trying a different route: instead of learning from data, they're building explicit physics models of absorption and using model-predictive control to plan each wipe in real time. That approach is more robust to new materials, but it's computationally heavy. The MPC loop needs to run at at least 50 hertz to keep up, and that takes a lot of onboard processing. Luna: It's a classic robotics trade-off: fast and adaptive but computationally expensive versus data-driven but brittle outside the training distribution. Lucas: Exactly. And what's really interesting is that this problem sits at the intersection of several challenges we've talked about before. The cloth manipulation from episode 57, the force control from episode 54, the tactile sensing from episode 52. Wiping a spill requires all of them working together. Luna: It's like a test case for general-purpose manipulation. If you can wipe a spill, you can probably do a lot of other contact-rich tasks. Lucas: Totally. And that's why I think it's worth paying attention to. By the way — this is exactly the kind of deep dive we love doing on this show, and we keep it free of ads because we think that makes the conversation better. If you find value in these episodes and want to support that choice, the link is buy me a coffee dot com slash fexingo. Luna: Yeah, it really helps keep the show going without any commercial breaks. Appreciate everyone who's contributed. Lucas: So back to the spill problem — one more detail I found fascinating. The Tokyo team discovered that the best strategy isn't to use a single cloth. They built a gripper that can automatically discard a saturated cloth and grab a fresh one from a dispenser. With that, their success rate jumped to 81 percent, because each cloth is used for only the first few seconds of absorption. Luna: So it's a hardware hack to avoid the control problem — just use multiple cloths. Lucas: Exactly. But that adds complexity and cost. And it doesn't solve the fundamental issue of adapting to the cloth state. It just sidesteps it. For a home robot that needs to work with whatever cloth is available, you still need the control solution. Luna: So where does that leave us? What's the timeline for a robot that can actually wipe a counter? Lucas: I'd say we're still a few years away from a reliable consumer product. The research is moving fast, but the combination of deformable objects and fluid dynamics is genuinely hard. My guess is we'll see a solution in industrial cleaning first — where the surface, cloth, and liquid are all known — and then it'll trickle down to home robots around 2028 or 2029. Luna: That seems reasonable. And in the meantime, I'll keep a paper towel handy. Lucas: Right. And when the robot does finally manage it, I think we'll look back at the spill-wiping problem as one of those milestones that suddenly made robots feel capable in a domestic setting. It's such a small thing, but it's a huge step.