Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Peel a Potato
Transcript
- Lucas: So, peeling a potato. It sounds trivial, but robotics engineers have been trying to automate it for decades. And they still can't get it right — not reliably, not at scale. Luna: Wait, we can put a robot on Mars, but we can't make one peel a potato? That's wild. What specifically makes it so hard? Lucas: Three main problems. First, potatoes are irregular — no two are exactly the same shape or size. Second, the material properties vary: some potatoes are waxier, some are more watery, the skin thickness differs. Third, you need precise force control — too much pressure and you gouge the flesh, too little and the peel won't come off. Luna: So it's not just about gripping. It's about adapting to an unpredictable object in real time. That's basically the holy grail of robotic manipulation. Lucas: Exactly. And we have some numbers to put this in perspective. A human can peel a potato in about 15 seconds. The best robotic attempts — from research labs — take several minutes, with a failure rate of 30 to 40 percent. And those are under controlled conditions. Luna: So what's the state of the art? Is anyone actually working on this? Lucas: There's a well-known study from Kyoto University in 2021. They tried to get a robot arm to peel apples — which are slightly more uniform than potatoes — using a standard peeler tool. They achieved a success rate of about 60 percent. And that was with a camera tracking the apple's shape and a force sensor in the wrist. Potatoes are even harder because they're knobbier. Luna: Sixty percent — so out of ten apples, four are ruined? That's not exactly ready for the supermarket shelf. Lucas: No, it's not. And the reason goes deep into robot design. Most industrial robot arms are built for repeatability — doing the exact same motion thousands of times. But peeling requires compliance — the ability to gently follow a contour. That's why soft robotics and impedance control are such active research areas. Luna: And this isn't just a fun kitchen problem. If you can peel a potato reliably, you can probably handle a lot of other delicate tasks — like deboning fish, or trimming meat. There's a huge economic incentive. Lucas: Yeah, the food processing industry is desperate for this. Labor shortages in agriculture and food prep are getting worse. If you could automate peeling, you'd change the economics of countless products — from frozen fries to canned vegetables. Luna: Let's dig into the mechanics a bit. What's the actual failure mode? Why does the robot mess up? Lucas: A couple of common ones. One is losing track of the potato's orientation — the peeler slips, or the potato rotates in the gripper. Another is applying too much force and cutting deep. Or too little force, so the peel just skims off a thin layer and then breaks. The robot's feedback loop has to adjust in milliseconds, and that's computationally expensive. Luna: So it's a real-time control problem. And the sensors matter a lot. Tactile sensors — like artificial skin — could give the robot a sense of how much pressure it's applying, but they're still not as good as human fingertips. Lucas: Right. Human skin has thousands of mechanoreceptors. A typical tactile sensor on a robot gripper has, say, a few dozen. So the robot is essentially working blind compared to us. Luna: And that's why machine learning is such a big part of the picture. You can train a neural network to map sensor data to grip adjustments, but you need huge amounts of data. And potatoes are diverse. Lucas: Which is exactly what some groups are doing. Researchers at MIT and Stanford have been using reinforcement learning to train robot arms to manipulate deformable objects — like a rope or a piece of cloth. Peeling is a similar kind of problem, but with the added difficulty of cutting. Luna: So we're not there yet, but the pieces are coming together. Better sensors, better algorithms, better hardware. It might take another decade. Lucas: I think that's a realistic estimate. If today's tech conversation gave you something usable — a new perspective on robotics, or just a good example of how hard the 'simple' things are — then you might want to consider supporting the show. Listener support is what keeps this podcast ad-free and focused on the actual engineering. You can buy me a coffee at buy me a coffee dot com slash fexingo. It really does help us keep digging into topics like this. Luna: Absolutely. And as we mentioned, the economic stakes are huge. If we can solve potato peeling, we can solve a lot of other manipulation problems in manufacturing and logistics. Lucas: So back to the present. One promising approach is using a combination of vision and force control with a soft gripper. A company called Soft Robotics Inc. makes grippers that can handle delicate items like fruit without damaging them. They've had success with picking apples and tomatoes. Peeling is the next frontier. Luna: But with potatoes, you also have the issue of removing the peel cleanly. You don't just want to rip it off; you want a smooth surface. That requires a cutting tool that follows the contour precisely. Lucas: Yes. Some researchers have tried using a rotating blade like a mandoline, but then you need to feed the potato in at a consistent angle. That's another control problem. Luna: So it's not just the gripper and the arm; it's the tool design too. A potato peeler is a simple device, but using it effectively requires a sophisticated understanding of force and geometry. Lucas: Exactly. And that's why I think peeling a potato is actually a fantastic benchmark for robotic dexterity. It's a single task that tests perception, planning, control, and tool use all at once. The industry should adopt it as a standard test. Luna: Speaking of benchmarks, there's the yale cmu berkeley object manipulation benchmark, but it doesn't include peeling. Maybe someone should add a 'potato challenge' to the robotics competitions. Lucas: That would be fun to watch. But seriously, progress is being made. I expect that within five years we'll see a robot that can peel a potato with 90 percent success in a lab. Then it's a matter of cost and speed for commercial use. Luna: And then we can finally automate the production of mashed potatoes. But what about other vegetables? Carrots? Cucumbers? Lucas: Carrots are easier because they're more uniform. Cucumbers are harder because they're more delicate. Each vegetable presents its own challenges. But the same principles apply. Luna: So to wrap up, here's a thought: the things we find easy — like peeling a potato — are actually incredibly complex. And understanding that complexity is humbling. It also tells us a lot about what human hands can do. Lucas: Absolutely. Every time a robot fails at something like this, it reminds us how remarkable our own biology is. And that's the real takeaway: we're not building better robots just to replace ourselves, but to understand ourselves better. Luna: Well said. And on that note, thanks for listening. See you next time.