Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Peel a Potato
Transcript
- Lucas: There is a YouTube channel run by the Robotics Institute at Carnegie Mellon that posts these one-minute clips of a robot arm trying to peel a potato. And it is strangely compelling. Luna: I've seen it. The arm just kind of hovers there, twitching, and then it stabs the potato and drags it across a peeler. It looks like a toddler who's been given a knife. Lucas: Exactly. And that is not an insult to Carnegie Mellon. That is the state of the art. We have robots that can perform open-heart surgery, assemble a transmission, even flip a burger. But peeling a potato — something a human can learn in about thirty seconds — is still a research-level problem. Luna: So what makes the potato so hard? I mean, it's not moving, it's not trying to escape. Lucas: The potato has three properties that break most industrial robot grippers. First, it's irregular. There is no fixed geometry. So you can't just program a set of coordinates and have the arm grab it the same way every time. Second, it's soft. The skin gives a little, but the flesh underneath is brittle — if you squeeze too hard, you bruise the potato. Too soft, and it slips. Luna: And third? Lucas: Third, the peel is a variable-stiffness surface. Near the tip, it's thin and flexible. Near the middle, it's thicker. A potato that's been sitting in a bin for two months has a different texture from one harvested last week. So the robot can't just learn one grip and one peeling stroke. It has to adapt in real time. Luna: And this is not just about potatoes. This is the same problem that makes it hard for a robot to pick up a tomato without crushing it, or to handle a raw egg, or to fold a towel. Lucas: Right. The potato is a benchmark. And the Carnegie Mellon group, led by a PhD student named Yifan Sun, published a paper last year that got a lot of attention. They used a very cheap force sensor — about two hundred dollars — mounted on the wrist of a Universal Robots UR5 arm. And they trained a neural network on about ten thousand peeling trajectories. Luna: Ten thousand trajectories. So the robot basically practiced peeling potatoes for days. Lucas: Weeks. The training took about three weeks of continuous runtime. And the result? The robot could peel a potato with about eighty percent success — meaning it got most of the skin off without mangling the potato — in about ninety seconds. Luna: Ninety seconds is slow. A human does it in maybe fifteen, and with near perfect yield. Lucas: But for a robot, that ninety seconds is a breakthrough. Three years ago, the same task had a success rate below twenty percent. So the field is moving. But it's moving at a pace that tells you something fundamental about how different robot intelligence is from human intelligence. Luna: Let's talk about why. Humans have about four million nerve endings in our fingertips. A typical robot gripper — even a fancy one — might have a dozen tactile sensors. So the feedback loop is just much coarser. Lucas: That's the hardware side. But there is also a software problem. Humans don't need ten thousand trajectories to learn to peel a potato. We watch someone do it once, maybe twice, and then we can generalise. A robot has to be shown thousands of examples, and even then, if you give it a different variety of potato — say a Yukon Gold instead of a Russet — it might fail completely. Luna: So the gripper technology is catching up. There are companies like Soft Robotics in Boston that make these pneumatic grippers that can conform to irregular shapes. And there are tactile sensors like the ones from GelSight, which use a camera and a gel pad to give high-resolution touch feedback. Lucas: Right. But the integration is still hard. You need a gripper that can provide enough force to hold the potato steady during peeling, but not so much that it damages the surface. And you need a control algorithm that can adjust that force on the fly, at millisecond timescales, based on the tactile data. That's a real-time control problem that most industrial robot controllers aren't designed for. Luna: And the peeling action itself — you have to apply a certain angle and pressure with the peeler, and you have to know where the peel ends and the potato begins. The robot doesn't have a concept of 'skin'. It's just looking for a change in force or colour. Lucas: That's a great point. The Carnegie Mellon team actually used a combination of force and vision. A camera above the workspace identified the potato's orientation, and the force sensor detected when the peeler had cut through the skin. But even with both modalities, the robot still occasionally gouged out a chunk of flesh or missed a strip of peel. Luna: So where does this leave commercial applications? Are there any companies using this kind of dexterous manipulation in production? Lucas: A few. In food processing, there are machines that can peel apples or potatoes at industrial scale, but they're not using general-purpose robot arms. They're custom-built machines with mechanical fingers and water jets. The flexibility to switch from potatoes to carrots to onions — that's still a research challenge. The same is true for handling soft fruits in packing plants. A robot that can pack strawberries without bruising them is the holy grail, and we're not there yet. Luna: The economic incentive is huge, though. Labor shortages in agriculture and food processing are getting worse. If you could build a robot that could handle a variety of soft, irregular objects with human-level dexterity, you'd have a billion-dollar product. Lucas: Absolutely. And that's why so many labs are working on it. But I think the potato peeling problem is a good reality check. It shows how far we have to go. We have self-driving cars that can navigate city streets, but we can't build a robot that can reliably peel a potato. Luna: It's humbling. And it's a reminder that the last inch of human capability — fine manipulation — is the hardest to automate. Lucas: Speaking of things that keep the lights on in these labs — listener support is a big part of how independent shows like this stay ad-free. If today's conversation gave you something useful, something you might mention to a friend over coffee, that sort of thing genuinely matters. A couple of dollars a month at buy me a coffee dot com slash fexingo — it adds up and it keeps us from having to run interruptive ads. Luna: Yeah, and it's a way to signal that this kind of deep dive — not clickbait, not hype — is worth doing. We appreciate everyone who's chipped in. Lucas: Okay, back to the potato. One of the more interesting follow-ups to the Carnegie Mellon work came from a group in Japan that tried a completely different approach. Instead of learning from demonstration, they used a physics simulator to generate millions of synthetic peeling trajectories, then transferred that policy to a real robot. Luna: sim to real transfer. That's a big area in robotics right now. How well did it work? Lucas: Surprisingly well. The robot could peel a potato with about seventy-five percent success after training only in simulation. But the researchers noted that the sim to real gap was still significant — the simulator couldn't perfectly model the friction of the potato skin or the deformation of the flesh. So the robot sometimes applied too much force and crushed the potato. Luna: That's the same problem: the physics of soft objects is just hard to simulate accurately. Lucas: Right. And that's the underlying theme of this whole episode. The potato is a stand-in for a class of problems that involve deformable objects — things that change shape when you touch them. That includes food, but also textiles, human tissue, rubber gaskets, foam packaging. Manufacturing and logistics are full of these objects, and robots still can't handle them well. Luna: So what's next? Is there a particular technology that could break the logjam? Lucas: A lot of people are betting on better tactile sensing. The GelSight sensors I mentioned earlier — they can provide a high-resolution map of contact pressure, almost like a fingerprint. If you combine that with a very fast control loop, you might be able to achieve the kind of adaptive grip that humans use without thinking. Luna: And there's also work on soft grippers that can vary their stiffness on the fly — like a balloon that can go from floppy to rigid. Lucas: Exactly. Researchers at Harvard have a gripper that uses a granular jamming mechanism. It's a bag filled with coffee grounds. When you apply a vacuum, the grounds lock together and the bag becomes stiff, conforming to the shape of the object. That's a clever solution, but it's still not fast enough for high-speed production lines. Luna: So we're looking at a combination of better hardware and better algorithms. The machine learning side has made huge strides in the last five years, but the hardware is still catching up. Lucas: I think that's the right takeaway. The potato peeling problem isn't going to be solved by a single breakthrough. It's going to be solved incrementally, as force sensors get cheaper and more sensitive, as grippers get more compliant, and as reinforcement learning algorithms get better at handling real-world physics. Luna: And until then, we'll keep watching those Carnegie Mellon YouTube videos and marveling at how hard a simple task can be. Lucas: Maybe that's the real value of the potato benchmark — it keeps the field honest. It reminds everyone that we haven't cracked manipulation yet. And that's a good thing to remember, especially when you see a press release about a robot that can fold laundry or cook a meal. Usually, those demos are heavily staged. Luna: Right, the lab conditions are perfect. The lighting is controlled, the objects are placed exactly the same way every time. The robot fails if you move the potato an inch to the left. Lucas: So the next time you peel a potato yourself, take a second to appreciate what your brain and your fingertips are doing automatically. It's a miracle of biological engineering that we've only begun to replicate.