Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Slice a Tomato
Transcript
- Lucas: Okay, Luna, I've got a new one for you. How hard can it be to slice a tomato? Luna: For me? Not hard at all. For a robot arm? I'm guessing this is going to be another episode in our ongoing series of 'things humans do effortlessly that robots completely fail at.' Lucas: Exactly right. And this one is a perfect storm. Tomato slicing combines almost every hard problem in robotics: deformable objects, variable friction, real-time force control, and precise tool handling. A chef does it in under a second. A robot? It's a research project. Luna: So where does it break down? Is it the blade, the tomato, or the arm? Lucas: All three. Let's start with the tomato itself. It's a viscoelastic object — it deforms under pressure, but not uniformly. The skin is tougher than the flesh, the seeds and gel inside have different densities, and as you cut, the geometry changes. A robot arm programmed with a fixed path and constant force will either not pierce the skin or will squish the tomato. Luna: So it needs to sense the resistance and adapt in real time. That's force control, right? Lucas: Right. Force control has been around for decades in industrial robots — but mostly for rigid tasks like grinding or deburring metal parts. Those applications assume the workpiece is stiff. A tomato is the opposite. It yields. And the robot has to modulate its downward force and cutting speed continuously, based on feedback from the knife. Luna: And the knife introduces another variable — friction. The blade has to slide through the tomato, but friction changes with the angle, the sharpness, and the juice. Lucas: Exactly. In 2023, researchers at MIT's CSAIL published a paper specifically on robotic tomato slicing. They used a six-axis arm with a force-torque sensor at the wrist, and they trained a model to predict the force required based on visual and tactile data. Even then, success rate was only around 70 percent for a single slice. And that's under lab conditions — controlled lighting, known tomato variety, consistent ripeness. Luna: Seventy percent in a lab. So in a real kitchen, it's probably much worse. What about the speed? How long did it take? Lucas: The MIT arm took about 15 to 20 seconds per slice. A human is under one second. So we're orders of magnitude off. And that's just one slice. To dice a tomato — multiple cuts in different orientations — you'd need to reposition the knife, account for the tomato rolling, and avoid the partially cut sections shifting. It's a combinatorial nightmare. Luna: This is why we don't have robot chefs yet. But there are industrial applications where cutting soft materials is critical. What about meat processing? Or fruit packing? Lucas: Great question. In meat processing, robots have been used for tasks like cutting chicken breasts or pork loins for years. But those are semi-structured — the meat is on a conveyor, it's roughly uniform, and the cuts are repetitive. Even then, yield loss due to inaccurate cuts is a major cost. Companies like Fanuc and ABB have dedicated food-grade robots with wash-down housings, but they still struggle with irregular shapes and variable density. Luna: So the tomato problem is really a subset of a bigger challenge: handling deformable, non-rigid objects. We've talked about this before with the raw egg episode, but the tomato adds a cutting tool into the mix. Lucas: Right. And that's the critical difference. With an egg, you're just gripping — you need to avoid crushing it. With a tomato, you're applying a controlled shear force through a blade. The interaction between the tool and the object is dynamic. And the robot has to manage not just the force, but the torque and the blade angle. Luna: I remember reading about Amazon's robotic depalletizing systems. They use suction cups and vision to grab boxes, but they avoid touching anything deformable. That's why Amazon Fresh still has humans handling produce. Lucas: Exactly. And that's a massive bottleneck. Grocery warehouses are one of the least automated sectors. According to a 2025 report from McKinsey, only about 15 percent of tasks in grocery fulfillment are automated, compared to 60 percent in general e-commerce. The reason? Deformable objects and fresh produce. Luna: So if you could reliably slice a tomato, you could probably automate a lot of food prep. That's a huge market. Lucas: It is. The global food robotics market is projected to hit 30 billion dollars by 2030. But the technological gaps are still wide. One promising approach is what's called 'tactile-driven cutting.' Instead of relying solely on vision, researchers are embedding tactile sensors in the knife itself — measuring pressure distribution along the blade. Luna: That's fascinating. So the knife becomes a sensor. You get real-time feedback on where the blade is relative to the tomato's internal structure. Lucas: Yes. A group at the University of Tokyo demonstrated a knife with 16 capacitive sensors along the edge, sampling at 1,000 hertz. They could detect the transition from skin to flesh within 2 milliseconds and adjust the cutting speed accordingly. Their success rate on slicing tomatoes was 92 percent — much better than the MIT approach. Luna: But that's still a research prototype. How far are we from a commercial system that a restaurant or a food processor could buy? Lucas: Probably 5 to 10 years. The sensors themselves are expensive, and the control algorithms need to be robust to different varieties, ripeness levels, and even ambient temperature. Plus, you need to combine it with vision to locate the tomato and decide where to cut. Integration is the hard part. Luna: And cleaning. Food-grade robots have to be washable. Sensors add complexity there too. Lucas: Good point. So the tomato is a benchmark problem. If you can solve it, you've solved a huge class of manipulation tasks. And it's not just food — think of surgical robotics, where you're cutting soft tissue. The same force control and tactile sensing challenges apply. Luna: Right. So the tomato slice is actually a microcosm of one of the most important open problems in robotics. And it's a good reminder of how much we take for granted in our own motor skills. Lucas: Exactly. Every time I slice a tomato at home, I think about all the real-time sensor fusion my brain is doing without me even noticing. It's humbling. Luna: You know, if today's episode gave you something useful — a new perspective on how hard robot manipulation really is — the way we keep this show ad-free is through listener support. You can buy us a coffee at buy me a coffee dot com slash fexingo. Every little bit helps us keep digging into these topics. Lucas: Yeah, and we really appreciate it. It's what lets us spend time on episodes like this one, rather than chasing sponsors. So if you're enjoying the series, that's the link. Luna: Alright, back to the tomato. Lucas, you mentioned the MIT CSAIL paper — what was their approach? Did they use any learning-based methods? Lucas: They did. They used a deep reinforcement learning framework where the robot was rewarded for completing a clean slice without squishing the tomato. The action space was the force applied and the cutting speed. But the simulation to reality gap was significant — the tomato in simulation behaved differently from real tomatoes. They had to do a lot of domain randomization to get decent transfer. Luna: So you train in simulation, then fine-tune on real tomatoes? That sounds data-hungry. Lucas: It is. And each tomato is a new data point. You can't reuse the same tomato for multiple cuts — you ruin it. So data collection is slow and expensive. That's another bottleneck. Luna: What about using a different tool? Like a serrated knife? Or a wire cutter? Lucas: Some researchers have tried serrated blades because they require less downward force — the teeth grip the skin. But the friction is more complex, and the cut quality is worse. For a nice, clean slice, you want a straight, sharp blade. Wire cutters work for soft cheeses, but for tomatoes, they tend to deform the shape too much. Luna: So it's back to the straight blade. And that means force control is the key. Lucas: Force control and also impedance control — managing the relationship between force and position. When you cut, you want the blade to move at a constant velocity, but the force needed varies. If you control force directly, the blade might speed up through soft spots and slow down at the skin. The ideal is a hybrid controller that switches between force and position modes depending on the feedback. Luna: That sounds complex to tune. Are there any off-the-shelf controllers that do that? Lucas: Some industrial robots offer 'force control' as an option — Fanuc has their Force Sensing function, ABB has Integrated Force Control. But these are designed for assembly tasks, not cutting. They assume the contact is rigid. For cutting, you need much higher bandwidth and sensitivity. So most researchers still build custom controllers. Luna: Is there a simpler approach? Like, what if you just use a very sharp knife and a light touch? Lucas: A sharper knife does reduce the force needed, but then the robot has to position the blade extremely accurately to avoid slipping. Slipping is dangerous — the knife could cut into the robot's own wrist or into a human nearby. So safety constraints add another layer. Luna: And that's why collaborative robots are usually slow and weak. They can't wield a sharp knife at high speed. Lucas: Exactly. Most collaborative robots have force limits to prevent injury. You can't use them for high-force tasks. So slicing a tomato with a cobot is actually safer, but you trade off speed. And speed is what makes the economics work in a commercial kitchen. Luna: So where does that leave us? A few years out from practical tomato-slicing robots, but with a clear research path. Lucas: I'd say so. And the progress in tactile sensing and force control is accelerating. Ten years ago, we couldn't even reliably grip a tomato. Now we're slicing them at 70 percent success. That's real progress, even if it sounds slow. Luna: And maybe by episode 60, we'll have a robot that can dice an onion without making you cry. Lucas: Don't hold your breath. Onions are even harder — they have layers that separate, and the tear gas is a whole other problem. Luna: Alright, so the humble tomato is a perfect test case for the next generation of robotic manipulation. Thanks for breaking it down, Lucas. Lucas: My pleasure. And if you're curious to see the MIT CSAIL paper, it's publicly available — just search for 'robotic tomato slicing' and you'll find it.