Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Wipe a Countertop
Transcript
- Lucas: So we've spent a lot of episodes on what robots can't pick up, what they can't fold, what they can't thread. But there's a task that sounds even simpler on paper, and it's one that almost every home and workplace needs done daily: wiping a countertop clean. Luna: Wiping. Like, with a cloth or a sponge. How hard can that be? Lucas: Harder than threading a moving needle, apparently. Last month at the IEEE International Conference on Robotics and Automation, there was a whole workshop dedicated to deformable object manipulation. And wiping came up as one of the 'grand challenges' because it combines so many unsolved problems at once. Luna: Such as? Lucas: Let's start with the obvious: you need to know where the surface is, and it's not always flat. Countertops can have seams, edges, slight curves. A robot arm with a rigid end effector—say, a standard gripper holding a sponge—has to maintain constant, even pressure across the entire stroke. If you press too hard, you smear the spill. If you press too lightly, you just move it around. Luna: And if the spill is something viscous, like maple syrup or ketchup, the force feedback changes as the cloth absorbs it. Lucas: Exactly. That's the second layer. Most industrial robots operate in position-control mode: 'move to these coordinates, with this speed.' But wiping is inherently force-controlled. You have to adapt in real time to the changing friction and compliance of the cloth, the liquid, and the surface. Researchers at the University of California, Berkeley did a study in 2024 where they tried to get a KUKA LBR iiwa—that's a collaborative robot with torque sensors in every joint—to wipe a dry whiteboard. Even with perfect force sensing, the robot left streaks because it couldn't correct for micro-variations in the board's flatness. Luna: So the robot doesn't know if it's actually removing the substance or just redistributing it. Lucas: Right. And that brings us to the third problem: detection. A human sees a spot and wipes until it's gone. A robot typically doesn't have vision that can distinguish 'clean' from 'dirty' at a fine enough resolution, especially if the spill is translucent or the same color as the surface. Some labs have tried using near-infrared cameras or UV markers, but that's not practical for a kitchen counter. Luna: I remember reading about a team at Carnegie Mellon that used a tactile sensor array on the back of a cloth—basically a 'touch camera' that could feel if the cloth was slippery or sticky. But the resolution was still coarse. Lucas: Yeah, that was the GelSight sensor work. The problem is scale: a GelSight sensor gives you high-resolution tactile data over an area about the size of a quarter. Wiping a countertop requires coverage of several square feet. You'd either need an array of those sensors—expensive and bulky—or you'd need to sweep the sensor across the whole surface, which is slow. Luna: So the state of the art in 2026 is that a robot can wipe a small, flat, dry surface if you pre-program the path and the force. But the moment you add a liquid, a crumb, a curved edge, or a change in the cloth's wetness, it falls apart. Lucas: That's a fair summary. And it's not just academic. There's a huge commercial demand for autonomous cleaning in hospitals, restaurants, and labs. Companies like Brain Corp and iRobot have made progress with floor mopping—but that's a flat, open surface, and the robot can use a passive mop that just drags along. Wiping a countertop is fundamentally different because it's a dexterous, contact-rich task with a highly deformable tool. Luna: What about the soft robotics approach? I've seen videos of a soft gripper that can wring out a sponge. Lucas: That's from MIT CSAIL, actually, from a paper published in early 2025. They built a soft, skin-like gripper with embedded pneumatic channels that can not only grasp but also squeeze and slide. In their demo, they had it wipe a small puddle of water off a glass table. It worked—but the conditions were tightly controlled: the table was perfectly flat, the water was dyed blue for visual tracking, and the robot had a pre-recorded trajectory. When they added maple syrup, the gripper got stuck because the syrup's tackiness exceeded the peeling force of the soft material. Luna: So the very compliance that makes it good for conforming to surfaces also makes it bad at handling sticky substances. Lucas: Precisely. That's the trade-off in soft robotics: you gain adaptability at the cost of precision and sometimes at the cost of release. In the same paper, they noted that the soft gripper could wipe a dry board clean of chalk dust, but it couldn't wipe a wet spill without leaving residue because the sponge itself became saturated and the gripper couldn't wring it out in place. Luna: So what's the most promising direction you've seen? Lucas: I think it's a hybrid approach: a rigid arm with force-torque sensing, combined with an end effector that has some passive compliance—like a spring-loaded wiper—plus a separate 'dirt detection' system using optical or thermal sensing. There's a group at Stanford that developed a 'smart squeegee' with a pressure-sensitive edge and a camera that looks at the streak pattern to adjust the angle. In their 2025 preprint, they achieved 95 percent removal of a coffee stain on a glass surface in a single pass, which is the best I've seen. Luna: But a squeegee isn't a cloth. Does it handle irregular surfaces? Lucas: No, it doesn't. That's the next frontier. The squeegee works on glass and smooth metal, but on a textured countertop—like butcher block or tile—it skips and leaves streaks. So the research community has basically concluded that there is no universal wiping solution. You have to specialize: a robot for glass, a robot for tile, a robot for spills, a robot for dust. And that's expensive. Luna: Which is why we still wipe our own counters. But it's interesting that this is where the frontier is—not in picking up a grain of rice, but in something as mundane as wiping. Lucas: Yeah, and that's the theme of this whole series: the tasks that seem trivial to us are often the hardest for robots because they involve continuous, adaptive contact with the environment. We've covered picking, folding, threading—but wiping is arguably the most practical. If you could automate cleaning a countertop, you'd have a product that every restaurant, every lab, every home would want. Luna: And we haven't even talked about the need to change the cloth when it gets dirty. That's a whole other can of worms. Lucas: Absolutely. Some labs are working on robotic cloth-changing using a separate gripper to peel the dirty cloth and snap on a clean one from a dispenser. But that adds mechanical complexity and another point of failure. So for now, the human hand remains the best tool for the job. Luna: You know, speaking of tools that are still the best—we've been doing this series for a while now, and one thing I really appreciate is that we've kept these episodes free from any sponsor interruptions. No ad reads, no 'this segment brought to you by.' Lucas: Yeah, that was a deliberate choice from the start. We wanted the conversation to be clean, no breaks. The only reason we can do that is because some listeners support the show directly. If you've gotten something out of these episodes—a new way to think about robotics, or just a good story—and you want to help keep it that way, the link is buy me a coffee dot com slash fexingo. Luna: It's a small gesture that makes a big difference. And it means we never have to tailor a segment around a product pitch. Anyway, back to the science. Lucas: So let's talk about what might finally crack this problem. One emerging idea is 'learning from demonstration'—having a human guide the robot arm through a wiping motion, and then having the robot generalize that motion to new surfaces and spills. A 2026 paper from the Robotics Institute at CMU used a technique called 'implicit force imitation' where the robot learns the force profile from human demonstrations, not just the trajectory. In their tests, the robot could wipe a countertop with a 90 percent reduction in residue compared to a purely geometric path. Luna: Is that being deployed anywhere? Lucas: Not yet. But the same group is working with a commercial cleaning equipment manufacturer to pilot it in a hospital environment, specifically for disinfecting bedside tables. The big hurdle is variability: each table has different clutter, different liquids, different contamination levels. The robot needs to adapt to each scenario in seconds, not minutes. Right now, the learning pipeline takes about a minute to process a new demonstration and generate a force profile. That's too slow for dynamic environments. Luna: So the bottleneck is compute, not hardware. Lucas: Largely, yes. The hardware exists—collaborative arms with good force sensing, soft grippers, smart cloths. But the real-time perception and decision-making to handle arbitrary spills is still beyond what we can do on an embedded processor. Most researchers run their algorithms on a laptop tethered to the robot, which isn't practical for a commercial product. Luna: What about using a foundation model trained on video of humans cleaning? Could that shortcut the learning? Lucas: It's being tried. A group at Google DeepMind published a paper last month—'RoboCleaner: Learning Wiping Policies from Internet Videos'—where they used a model trained on YouTube videos of people cleaning to output joint torques for a robot arm. It worked on simple spills—water, juice—but failed on sticky substances because the model had never seen the tactile feedback of a cloth getting gummed up. So the vision-language model can tell you 'wipe in a circular motion,' but it can't feel when to increase pressure. Luna: So we're back to the tactile sensing gap. It's always the tactile sensing. Lucas: It really is. For decades, robotics focused on vision because cameras got cheap and fast. But touch is a whole different sensor modality, and we're still in the early days of high-resolution, affordable tactile skins. Once we have a 'touch camera' that can cover a whole palm with hundreds of taxels at a reasonable cost, wiping will get a lot easier. Luna: And maybe then we'll finally have a robot that can clean up after a dinner party. Lucas: That's the dream. But for now, I'll still be the one reaching for the sponge. Thanks for listening—next episode, we're going to look at something even more slippery: picking up a single ice cube.