Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Can't Polish a Surface to a Mirror Finish
Transcript
- Lucas: I've been looking into a task that sounds simple but is actually one of the toughest unsolved problems in industrial robotics: polishing a metal surface to a mirror finish. Luna: Polishing? Like, the same kind of finish you'd see on a high-end faucet or a car's chrome trim? Lucas: Exactly. That kind of surface requires consistent force, precise path control, and the ability to adapt as material is removed. And right now, most factories still rely on skilled human workers for that final step. Robots can do rough grinding and deburring, but achieving that glossy, flawless mirror finish is a different story. Luna: So what makes it so hard? Is it the force sensing? Lucas: That's part of it. But let's start with a specific number: a human polisher can maintain contact force within about one newton while moving a tool across a curved surface. Commercial force-torque sensors on robot wrists are accurate, but they struggle with the high-frequency vibrations and the changing contact geometry during polishing. Also, the tool — often a buffing wheel or abrasive belt — wears down over time, changing the contact area. Luna: One newton — that's impressive. And it's exactly the kind of detail that makes you appreciate the craft. You know, it's conversations like this that remind me why we keep the show ad-free. A few listeners help make that possible through buy me a coffee dot com slash fexingo, and that support lets us dig into topics like this without any commercial interruptions. Lucas: Yeah, we really value that. And it keeps the podcast exactly the way we want it — focused on the technology and the people behind it. So, back to polishing: the challenge isn't just force, it's also the path planning. Luna: Path planning? You mean the robot has to follow the contour of the part? Lucas: Exactly. A typical car door outer panel, for example, has compound curves. The robot's tool needs to maintain constant overlap and speed across that whole surface. Any deviation can leave swirl marks or uneven gloss. In practice, that means generating a toolpath that matches the exact CAD model — but production parts have small variations from stamping or casting. The robot doesn't know where those variations are. Luna: So the robot is blind to the actual shape of the part? Lucas: Basically. Some setups use laser scanners to map the surface before polishing, but then you're adding cycle time. And even then, the robot can't adjust in real-time the way a human does. A polisher sees the shine building up and backs off if it's too aggressive, or goes back over a dull spot. That kind of closed-loop visual feedback is still very hard to replicate. Luna: Do any robots actually succeed in a lab setting? Lucas: Yes, but with significant caveats. Researchers at the University of Stuttgart, for instance, demonstrated a robot that could polish a flat aluminum plate to near-mirror finish using force control and a learning algorithm. But they had to run multiple passes, and the process took several minutes — not competitive with a human. And on curved parts, the performance dropped. Luna: What about industry? Are there commercial polishing robots? Lucas: Companies like FANUC and ABB sell robotic polishing cells, typically for deburring or satin finishes. For mirror finish, they often still require manual touch-up. The problem is repeatability: a robot might do the same stroke perfectly a hundred times, but if the tool wears or the part varies, the result drifts. In a high-volume production line, that drift is unacceptable. Luna: So the human's adaptability is the key? Lucas: Yes. Humans use a combination of visual inspection, audio feedback — the sound of the tool changes with pressure — and tactile feeling. They can even detect surface defects by touch. Replicating that with sensors is an active area of research. Some labs are working with tactile skins that measure shear forces and vibrations, but they're not yet robust enough for factory floors. Luna: What about machine learning? Could a robot learn to polish from demonstrations? Lucas: There's promising work in that direction. At the Max Planck Institute, researchers recorded human polishing motions and used them to train a neural network to output force and trajectory commands. The robot could then match the human's finish on simple geometries. But it struggled with generalization — if the part shape changed even slightly, the model needed retraining. Luna: So we're still a long way from a fully autonomous polishing robot? Lucas: I'd say for true mirror finish, yes. But there's progress in related areas. For example, in aerospace, robots now do automated blending and smoothing of composite surfaces, but those are lower finish requirements. The gap between a satin finish and a mirror finish is surprisingly large — it's the difference between acceptable and perfect reflection. Luna: And perfect reflection is what sells high-end products. Lucas: Exactly. A luxury car buyer will notice a single swirl mark under showroom lighting. So automakers are reluctant to hand off that task to a robot until it's proven flawless. The cost of a rework or scrap part is much higher than the labor cost for a skilled polisher. Luna: Are there any new sensor technologies that might tip the scales? Lucas: One exciting development is high-speed optical coherence tomography — basically a 3D surface scanner that can measure roughness in microns at line rate. If you pair that with a fast force controller, you could theoretically have a robot that inspects and adjusts in real-time. But the processing power and latency are still challenges. And the sensor itself is expensive. Luna: So the economics aren't there yet for most applications? Lucas: Right. For now, the business case only makes sense for high-volume, simple-geometry parts. Think polishing stainless steel cookware — where the shape is a round bowl. Some factories have automated that. But for complex shapes with mixed materials, like a chrome-plated plastic trim, human hands are still the standard. Luna: It's fascinating that something so old — polishing — is still a frontier for robotics. Lucas: It really is. And it highlights a broader lesson: the tasks we think of as 'simple' often involve a rich stream of sensory feedback and motor adaptation that we don't even notice. Replicating that in a machine requires solving some of the hardest problems in control, perception, and learning. Luna: So what's the next step? Will we see a breakthrough in the next five years? Lucas: I think yes, but incrementally. Better force sensors, faster vision, and more robust learning algorithms will gradually eat away at the gap. I wouldn't be surprised if by 2030, a robot can polish a car door to a mirror finish in a fraction of the time a human takes, but only for a specific model. Generalization across parts will take longer. Luna: And that's the story of industrial robotics — solving one geometry at a time. Lucas: Exactly. And that's what keeps this field so interesting. Every time we think we've cracked manipulation, a new surface or material reminds us how much we still have to learn.