Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Struggle to Wipe a Counter
Transcript
- Lucas: Alright, I want to start with something that sounds absurdly simple: wiping a counter. Luna: Wiping a counter. Like, with a sponge? Lucas: Exactly. And if you think about it, a human can do it without thinking — you just grab a sponge, move it in a rough S-curve, and somehow you apply just enough pressure to pick up crumbs or sticky residue without smearing it. But for a robot arm, this is a genuine research problem. Luna: I've seen videos of those cleaning robots that just drive around in malls. They seem to do alright. Lucas: Those are floor scrubbers — big, flat pads with a lot of surface area, and they operate on uniform flat floors. A countertop wipe is different: you've got a small sponge, you're working on a surface that might have a slight lip or a non-smooth area, and you're trying to clean a specific spot, not a whole area. Plus, the robot has to deal with things like crumbs that move when you touch them. Luna: So the sponge itself is the sensor and the actuator at the same time? Lucas: That's exactly the framing. The robot has to feel the contact. Most robot arms rely on torque sensors in their joints to estimate force, but wiping is a dynamic contact task — you're sliding across the surface, so the friction changes constantly. If you press too hard, you'll smear the sticky stuff or even push the sponge into a gap. Too light, and you just push the crumbs around. Luna: I remember reading about a paper from Carnegie Mellon — maybe from 2023 — where they specifically studied robot wiping. They had a robot arm with a force-torque sensor at the wrist, and it had to wipe a whiteboard clean of marker. Lucas: Right, that's the famous 'whiteboard wiping' study. They found that the optimal strategy wasn't a simple back and forth. The robot had to learn a spiral pattern that overlapped its previous strokes by a certain percentage, and it had to adjust pressure based on the marker residue — which is nearly invisible to a camera. So they used the force sensor to infer where the residue was, because the friction coefficient changes when you're passing over a drawn-on area. Luna: So the robot is essentially feeling the dirt. Lucas: Exactly. And that's where it gets interesting. The paper showed that if you only use vision, the robot misses about 30 percent of the residue. But with force feedback, it can detect the change in friction and go back over those spots. That's a huge improvement — and it hints at how robots might handle more complex cleaning tasks in the future. Luna: And this is why your Roomba can't handle a spill. It just drives over it. Lucas: Precisely. A spill is a localised, non-uniform mess that requires directed pressure and a specific motion pattern. Roomba is a coverage machine — it doesn't know it's cleaning a spill. It just hopes to get lucky. But a real wiping robot needs to sense the spill, decide on a strategy, and then execute a wiping motion that's compliant enough to not just smear the liquid further. Luna: And then there's the issue of the sponge itself. A sponge is deformable, so the contact area changes depending on how much you press. Lucas: Yes, and that's a nightmare for modeling. Most robot control assumes rigid or known deformable objects. A sponge is neither. The stiffness varies with saturation, the shape changes, and the friction is highly nonlinear. So researchers have started using learning-based approaches — training a neural network to predict the force response from a given motion input. One lab at MIT used a dataset of thousands of real wiping trajectories, collected with a human teleoperating a robot arm, to train a model that could then perform the task autonomously. Luna: How close are we to having a robot that can actually wipe a kitchen counter in a home? Lucas: In a controlled environment, we're already there. Boston Dynamics showed a Spot robot with an arm that can wipe a whiteboard pretty effectively. And there's the Stretch robot from Hello Robot — a mobile manipulator with a soft gripper and a force sensor — that can do basic surface wiping. But in a real home with variable surfaces, unknown spill compositions, and changing lighting? We're probably three to five years away from a reliable product. The challenge isn't just the wiping motion; it's the perception and decision-making upstream. Luna: Honestly, if today's tech conversation gave you something usable — maybe a new appreciation for what your own hands can do — that's the link. Support the show at buy me a coffee dot com slash fexingo. Lucas: Yeah, listener support is what keeps this show ad-free and focused on the details. It's a small ask, and it makes a difference. Luna: And speaking of hands — let's talk about what makes human hands so good at this. Lucas: Great pivot. Human hands have thousands of tactile sensors in the skin — mechanoreceptors that detect pressure, shear, and vibration. Plus, you have proprioception — you know where your hand is in space without looking. And you have years of practice. When you wipe a counter, you're constantly making micro-adjustments: you see a crumb, you increase local pressure; you feel the sponge snag on a rough spot, you change the angle. A robot has to do all of that with maybe six torque sensors and a single camera. Luna: So it's not just about better algorithms. It's about better hardware. Lucas: Exactly. The tactile sensing is coming along. Companies like GelSight make high-resolution tactile sensors that can measure surface texture and slip. But integrating them into a practical cleaning tool — one that's waterproof, durable, and cheap enough for a home robot — is still a work in progress. And the wiping task itself is a great benchmark for the field, because it combines perception, control, and adaptation in a way that's directly applicable to many other manipulation tasks. Luna: Such as? Lucas: Sanding, polishing, painting — any surface finishing task in manufacturing. Even tasks like spreading adhesive or smoothing concrete. If you can master wiping, you can master a whole class of surface interaction tasks. That's why companies like Fanuc and ABB are actively funding research in this area. They want to automate quality control and finishing steps that currently require human touch. Luna: I was at a trade show last year where a startup was demoing a robot that could polish a car door panel. It used a force-controlled end effector and a vision system to detect imperfections. It was surprisingly good — but it took hours to program for one specific door shape. Lucas: That's the trade-off right now. You can get excellent performance on a specific task in a fixed environment. But generalising across different surfaces, different contaminants, different geometries — that's where the research frontier is. The Carnegie Mellon group I mentioned earlier is now working on a method called 'implicit wiping' where the robot learns a policy that works across multiple surface materials — wood, granite, laminate — by using a latent representation of the surface properties, inferred from the force feedback during the first few strokes. Luna: So the robot adapts as it goes, like a human would. Lucas: Exactly. And that's the key takeaway. The wiping problem isn't just about moving a sponge. It's about closed-loop adaptation — sensing the state of the surface and the tool, and adjusting in real time. That's a fundamental capability that will unlock a lot of other applications. Luna: One thing I wonder about: how do you evaluate success for a robot wipe? Is it visual inspection? A clean swipe on a white cloth? Lucas: Great question. In the lab, they use pre-weighed contaminant — like a known mass of peanut butter — and after wiping, they measure the residue left behind. Some groups use fluorescence: they add a uv detectable dye to the spill, and then use a UV camera to quantify how much remains. That gives them a precise metric. In the real world, the standard is 'clean enough for the next use,' which is subjective. But for research, they need objective measures. Luna: I also heard about a project from the University of Tokyo where they used a deformable sponge with embedded force sensors. So the sponge itself becomes the sensor. Lucas: Yes, that's a really clever approach. They embedded capacitive sensors inside a silicone sponge — essentially a soft, deformable tactile array. The robot could then measure local pressure distribution across the sponge surface, which gives much richer feedback than a single wrist sensor. They showed that with this setup, the robot could wipe a surface with variable topography — like a textured silicone mat — and maintain consistent cleaning performance. Luna: So the future might be smart sponges. Lucas: Smart sponges, smart surfaces — or maybe just better control algorithms that can work with the sensors we already have. There's a lot of low-hanging fruit in combining existing force-torque sensors with learning-based control. The hardware is getting cheaper and more reliable. It's really the software and the understanding of the physics that's the bottleneck. Luna: And the fact that spills are unpredictable. You never know exactly what you're wiping up. Lucas: Right. And that's the final piece of the puzzle — the robot needs to be able to handle novel situations. A drop of water behaves differently than a smear of grease. A crumb is different from a sticky spill. The robot has to identify the type of mess and adapt its strategy accordingly. Some groups are using vision-language models to do this — they show the robot a picture of the spill, and it generates a text description like 'sticky liquid spill on granite,' which then gets fed into the control system. Luna: That's actually pretty wild. So the robot is using a large language model to decide how to wipe? Lucas: In a sense, yes. The language model provides a high-level plan — maybe 'use a circular motion with medium pressure' — and then a low-level controller executes it. But the loop isn't closed yet. The robot can't tell if the plan is working until after the fact. That's where the force feedback comes back in. So you have this hierarchy: language model for strategy, force feedback for execution, and vision for initial classification. Luna: It's like a robot butler with a brain, eyes, and fingers all working together. Lucas: Exactly. And we're not there yet as a product, but the building blocks are coming together. I'd say within the next decade, you'll see a home robot that can handle a spill on your kitchen counter reliably. Not perfectly, but well enough to be useful. Luna: And maybe it'll also be able to fold towels and tie knots while it's at it. Lucas: Let's not get ahead of ourselves. One miracle at a time.