Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Wipe a Whiteboard Clean
Transcript
- Lucas: So, quick question: have you ever watched a robot arm try to wipe a whiteboard clean? Luna: Can't say I have. But I'm guessing it's not pretty. Lucas: It is genuinely bad. I watched a video from a lab at CMU a few weeks ago — they set up a UR5e with a standard dry-erase eraser, and the robot basically smeared ink into abstract art. It left ghost marks everywhere. And this is a twenty-thousand-dollar arm. Luna: Okay, so it's not a gripping problem, it's a wiping problem. That's a different class of manipulation. Lucas: Exactly. And that's the angle for today. We've done a lot of episodes on why robot hands can't grasp things — eggs, slippery soap, coffee cups. But wiping a surface is a whole other challenge. It's not about clamping force; it's about maintaining a precise, consistent contact force across a moving tool. Luna: And the tool is the eraser, which is compliant, right? The foam compresses. Lucas: Right. So the robot has to press hard enough to erase, but not so hard it stalls the motors or damages the board. And the squeeze — the amount of compression on the eraser — changes as it moves across the surface because your arm's geometry changes. Gravity, joint angles, all that. Luna: So we're talking about force control, not just position control. Lucas: Exactly. Most industrial robot arms are position-controlled. You tell them 'go to this X-Y-Z coordinate with this orientation,' and they get there with insane repeatability — like plus or minus 0.02 millimeters. But they're stiff. If you try to press a compliant eraser against a hard surface, the arm doesn't naturally modulate force. It either over-presses or under-presses. Luna: And the whiteboard eraser is foam, so it's nonlinear. The force-displacement relationship is not simple. Lucas: The CMU team — this was a paper at ICRA last year — they instrumented the eraser with a six-axis force torque sensor between the robot wrist and the eraser holder. So they could measure exactly how much force was applied in each direction. And then they ran hundreds of wiping trials at different speeds and forces. Luna: What did they find? Lucas: That the sweet spot for a standard dry-erase marker on a glossy whiteboard is about 4 to 6 Newtons of normal force — that's roughly the weight of a medium apple. And the wiping speed should be around 200 millimeters per second. Too fast and you skip, too slow and you smear. And the angle of the eraser matters — they found a 15 degree tilt off vertical was optimal. Luna: So they basically mapped the entire parameter space. But did they get the robot to actually wipe cleanly? Lucas: They did — under lab conditions, with a clean eraser, after calibrating the force sensor. But here's the catch: the robot had to do a 'priming' pass where it pushed the eraser against the board for a second to let the foam compress, and then it started the wiping motion. Without that initial compression, the force control oscillated. Luna: Right, because the foam has hysteresis. The first time you press, it's stiffer than after you've compressed it a few times. Lucas: Exactly. And that's the kind of subtle physical effect that makes these tasks hard. The robot can't just rely on a model of the eraser — it has to adapt in real time. There was another paper from Stanford around the same time that tackled this with a learning-based approach. They used a vision system to detect residual ink after each pass, and then adjusted the force for the next pass. Luna: So closed-loop control with visual feedback. That seems like the obvious solution, but it's computationally heavy. Lucas: It is, especially if you want real-time updates. The Stanford group used a camera mounted on the wrist, looking at the board from about 20 centimeters away. They used a neural network to classify the 'ghostiness' of the marks — how faint the remaining ink was — and then fed that into a PI controller that adjusted the normal force. It worked, but only at about 10 hertz. So the robot would wipe a strip, pause, look, adjust, wipe again. Not exactly fluid. Luna: But if you think about it, a human doesn't look at the board after every single pass either. We feel the resistance, we hear the squeak, we adjust continuously. Lucas: That's the real frontier — multi-modal feedback. Humans use force, tactile, auditory, and visual cues all at once. Robots today typically use one, maybe two. The CMU paper showed that even with a perfect force-torque sensor, you still get failures when the ink has been sitting on the board for a while — dried ink requires about 30% more force to remove. Luna: So the robot would need to know how old the marks are, or at least detect the adhesion level. Lucas: Right. And that's not something you can infer from force alone. You'd need a vision system to estimate ink age based on color or reflectivity. It's a cascade of sensing challenges. Luna: This episode actually made me think about something. If today's conversation gave you something usable — maybe you work in robotics or just enjoy understanding why your Roomba can't wipe a counter — the way these stay ad-free is listener support. You can find it at buy me a coffee dot com slash fexingo. It's a low-key way to keep the show going. Lucas: Yeah, exactly. It really helps. And back to the whiteboard — the other big challenge is that whiteboards aren't perfectly flat. They have slight curvature, and over time they get scratched. The CMU team tested on a brand new board and a board that had been used for three years. On the old board, the optimal force parameters shifted by about 15 percent. Luna: So the robot would need to adapt to surface wear as well. That's a lot of variables. Lucas: It's a great example of why 'simple' manipulation tasks are still unsolved. We have robots that can assemble car doors with sub-millimeter precision, but wiping a whiteboard clean in an arbitrary office environment? That's a research problem. Luna: Is there any commercial application? Like, are there companies trying to build a whiteboard-cleaning robot? Lucas: There are a few. One startup called BoardBot — I think they're out of Berlin — they have a device that mounts on the top of a whiteboard and moves a horizontal bar across. It uses a constant force spring, so it's purely mechanical. But it only works on boards that are exactly the same width, and you have to manually install it. No computer vision, no adaptation. Luna: So it's a niche product, not a general solution. Lucas: Right. For a general-purpose robot to do it, you need robust force control, surface sensing, and ink detection. And that's still at the lab stage. But the CMU work did get picked up by a company that makes collaborative robots — they're exploring a software module for wiping tasks in cleanroom environments. Luna: Cleanrooms, because they need to wipe down surfaces without leaving residue. Lucas: Exactly. In semiconductor fabrication, you have to clean wafers and surfaces with precise force to avoid contamination. So the whiteboard problem is actually a proxy for a high-value industrial task. Luna: That's a good point. So the research isn't just about whiteboards, it's about understanding contact mechanics for compliant tools. Lucas: Right. And that's why I think this area will see rapid progress in the next few years. We're getting better sensors, cheaper force-torque sensors, and better algorithms for real-time adaptation. I wouldn't be surprised if by 2030, a robot can wipe a whiteboard as well as a human. Luna: But will it do it with the same satisfying squeak? Lucas: Maybe they'll program that in as a feature.