Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Struggle to Cut a Watermelon
Transcript
- Lucas: So, we've spent over twenty episodes now on why robot arms can't handle a whole catalogue of everyday objects — slippery fish, chocolate truffles, wet bars of soap. Today I want to talk about something that sounds almost trivial but actually exposes a deep limitation in how robots perceive and act on the world. Luna: Let me guess — it's something seasonal? A watermelon, maybe? Lucas: Right on cue. Watermelon. Specifically, cutting one open. Think about what happens when you bring a knife down on a watermelon. You start on the rind, which is tough and fibrous. Then you hit the flesh, which is soft and brittle in a completely different way. Then you get near the bottom rind and you need to stop before you cut through the cutting board. For a human, it's a single fluid motion with subtle force adjustments. For a robot, it's a nightmare of non-linear mechanics. Luna: And it's not just about the varying textures. There's also the shape — a watermelon is round and unstable. If you push too hard, it rolls. Lucas: Exactly. So you've got three interlocking problems: geometric uncertainty, material heterogeneity, and the need for real-time force modulation. Most industrial robots operate on a stiff position-control loop. They're told where to be, and they go there, and if the resistance exceeds a threshold, they either stall or break something. That works fine for picking up a metal bracket and placing it in a jig. It fails when the object's properties change mid-motion. Luna: There was a really nice paper from the DLR — German Aerospace Center — a few years ago on their Agile Justin robot. They showed that a robot can cut a soft material like a tomato, but only if it combines force feedback with a very compliant control scheme. Lucas: Right, and a watermelon is a tougher test because it's a composite. The rind has a fracture toughness of about 2500 joules per square meter, while the flesh is an order of magnitude lower. So the knife encounters a sudden drop in resistance after about five millimeters. If the robot doesn't sense that transition and adjust its velocity, the blade will plunge through and the whole thing can split catastrophically. Luna: Which explains why most robotic cutting demonstrations you see online are carefully staged — the object is clamped, the blade path is pre-planned, and the material is uniform. Show me a robot that can cut a watermelon on a moving conveyor belt, and I'll show you a system that's actually ready for a commercial kitchen. Lucas: And that's the million-dollar application. Food processing is a huge industry — roughly two trillion dollars globally by some estimates — and it's still heavily manual. One reason is that fruits and vegetables are incredibly variable. A watermelon today might be harder or softer than one tomorrow, depending on ripeness, variety, storage conditions. Traditional robots can't adapt on the fly. Luna: So what would a watermelon-capable robot need? Better sensors, obviously, but also a fundamentally different control architecture. Lucas: Yes. The key is impedance control, which is a fancy way of saying the robot doesn't just command position — it commands a relationship between force and motion. It behaves like a spring with a tunable stiffness. When cutting through the rind, the robot can be stiff. When it hits the flesh, it needs to become compliant, reducing its effective stiffness so it doesn't overshoot. And it has to do that transition in milliseconds, based on tactile feedback. Luna: And tactile feedback itself is still a research frontier. We've got force-torque sensors at the wrist, but that's like sensing everything through your shoulder. You really want sensors IN the fingertips, or in this case, along the blade. Lucas: There's some promising work on distributed tactile sensors — thin films with arrays of capacitive or piezoresistive cells that can measure pressure distribution at high resolution. The challenge is robustness. A sensor that works on a bench in a lab gets destroyed after a few cycles in a wet, acidic food environment. Luna: And cost. That's the other barrier. A six-axis force-torque sensor from a company like ATI can cost several thousand dollars. For a food processing line with dozens of robots, that adds up fast. Lucas: So the engineering trade-off is clear: you can make a robot that cuts a watermelon with high reliability, but it'll cost more than a human worker doing the same job. The breakthrough will come when sensor and control costs drop to the point where the robot's speed and hygiene justify the upfront investment. Luna: I want to come back to something you said earlier about the geometry. Even if you solve the sensing and control, you still need to figure out where to cut. A watermelon isn't a perfect sphere, and the stem scar gives you a hint about orientation, but a robot would need a vision system that can estimate the center of mass and plan a cut that minimizes splitting. Lucas: Right. And that's a classic perception problem. Depth cameras work fine for boxy objects, but specular surfaces — shiny, curved, with high color contrast — are hard. There's a reason the standard test object in computer vision is a cube with AprilTags on it. A watermelon is the opposite of that. Luna: So we're looking at a system that must integrate: 3D vision for initial pose estimation, force-controlled cutting with adaptive impedance, tactile sensing for real-time feedback, and a model of the material's fracture mechanics. That's a lot of moving parts. Lucas: And yet, there are companies trying. One that comes to mind is a Japanese startup called Mujin — they focus on bin picking and depalletizing for logistics, but they've also demonstrated some impressive handling of irregular objects. Another is a us based company called Soft Robotics, which makes grippers that use pneumatic air chambers to conform to objects. They've done demos with fruit, but not cutting yet. Luna: And then there's the whole approach of learning from demonstration. Instead of programming the cutting trajectory explicitly, you have a human guide the robot through the motion, and the robot learns a policy that generalizes to new watermelons. That's been successful for some assembly tasks, but for cutting, the force profiles change a lot with each fruit. Lucas: That's where simulation can help. If you can build a realistic physics model of a watermelon — with variable rind thickness and flesh stiffness — you can train a policy in simulation and then transfer it to the real robot. That's the sim to real paradigm that's been so successful in manipulation research lately. Luna: But the sim to real gap is still significant for deformable objects. A finite element model of a watermelon that runs fast enough for reinforcement learning is going to be approximate. And if your policy is trained on approximations, it might not handle a real fruit that has a bruise or a weird shape. Lucas: All true. And here's the thing — I don't think the solution is going to come from one breakthrough. It'll be a combination of incremental improvements: cheaper tactile sensors, better simulation, more robust impedance controllers, and smarter vision. Each step gets us a little closer to the point where cutting a watermelon becomes a solved problem. Luna: Which raises a bigger question: once we CAN cut a watermelon reliably, what does that unlock? I mean, if you can cut a watermelon, you can probably cut most fruits and vegetables. That would be huge for the food industry. Lucas: Absolutely. It's a gateway capability. If you can handle the worst-case composite food item, then softer fruits like melons, tomatoes, even avocados become straightforward. And that means robotic systems could take over tasks that are currently done by human workers in repetitive, high-volume settings — think salad preparation, fruit packing, ready-meal assembly. Luna: There's also a safety angle. Knife injuries are common in commercial kitchens. A robot that can cut safely, with compliant control and fast stopping, could reduce that risk. Lucas: Right. And that's actually where I think the case for investment gets strongest. It's not just about replacing labor — it's about augmenting it. A robot that can handle the dangerous, dull, or dirty parts of food prep frees up human workers for tasks that require judgment and creativity. Luna: That's a nice framing. And it ties into something we've talked about before — the idea that robotics isn't about making human workers obsolete, but about making them more valuable. Lucas: Exactly. And speaking of value, this show is completely ad-free, and that's only possible because of listeners who support it directly. If you've gotten something useful from today's deep dive into watermelon cutting, and you'd like to help keep the episodes coming without commercial interruption, you can support the show at buy me a coffee dot com slash fexingo. Luna: Yeah, it's a small gesture that makes a big difference for us. Every bit helps us keep the production independent and focused on the topics that matter. Lucas: So, back to the cutting board. One more thing I want to touch on: the role of compliance in the tool itself. Most robotic cutting uses a rigid blade attached to a rigid wrist. But there's research showing that a spring-loaded blade — essentially a knife with a built-in suspension — can handle the rind to flesh transition much more gracefully without active control. Luna: That's clever. It's like passive compliance — you let the mechanics do some of the work that would otherwise require sensors and software. Lucas: Right. And it's a reminder that not every problem needs a smarter algorithm. Sometimes the right mechanical design can make the control problem trivial. There's a beautiful example from the animal kingdom: the mantis shrimp uses a spring-loaded appendage to deliver a punch that can break crab shells. It's a purely mechanical energy storage system. Luna: So maybe the future of robotic cutting is part mechanical innovation, part sensing, part control. A hybrid approach. Lucas: That's exactly right. And I think that's the takeaway for today: cutting a watermelon seems simple, but it's a microcosm of the entire manipulation challenge. It forces us to integrate perception, control, mechanics, and materials science. Solve it, and you've made a real dent in the dexterity problem. Luna: Well, I'm not going to test my robot on a watermelon just yet. But I'll be watching for the papers that come out of this area. Thanks, Lucas. Lucas: Thanks, Luna. And thanks to our listeners. Next time, we're going to look at something that sounds even simpler: why robot arms can't scoop ice cream. See you then.