Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Chop an Onion
Transcript
- Lucas: So I tried to teach my kitchen robot to chop an onion this weekend. Spoiler: the onion won. Luna: Wait, you have a kitchen robot? And you thought it could handle an onion? Lucas: Look, it's a Fanuc LR Mate 200iD I borrowed from a lab. Welds car doors perfectly. I figured an onion would be trivial. Turns out, the onion is a nightmare for force feedback. Luna: Is it the layers? The way the hardness changes from outer skin to inner flesh? Lucas: Exactly. You've thought about this more than most people. An onion is essentially a layered viscoelastic structure. The outer papery skin is stiff and brittle, then you hit a soft fleshy layer, then another, and each layer has a different water content. When a robot arm applies a downward force to chop, the blade hits a sudden change in resistance, and the controller either overcompensates — slamming the blade through — or undercompensates, stopping mid-chop. Luna: So the robot doesn't know how hard to push? Lucas: Right. Most industrial robots use impedance control: they model the expected force based on position. But an onion is irregular. The geometry changes from bulb to bulb, even inside the same onion. And the knife angle matters — a slight tilt changes the contact area. The robot can't 'feel' that the way a human does through skin and muscle spindles. Luna: I remember a paper from MIT CSAIL a couple years ago — they trained a model on over a thousand onion slices. Did that go anywhere? Lucas: Good memory. That was from 2024, I think. They used a UR5e with a custom force-torque sensor and collected a dataset of 1,432 onion slices. They trained a deep neural net to predict the optimal cutting force based on the real-time force profile. In the lab, it worked on about 85 percent of onions. But when they tried it on a different batch from a different farm — different water content — performance dropped to 60 percent. Luna: So the model overfitted to the training onions. Lucas: Exactly. And that's the core problem. Onions are biological — they have natural variation. The robot can't generalize from one batch to another without retraining. And retraining requires collecting new data, which means actually chopping a lot of onions. It's a chicken and egg problem. Luna: Is there any research on using soft grippers or compliant blades to handle the variability passively? Lucas: Yes. Some groups at Stanford and ETH Zurich are working on passive compliance. Instead of actively controlling force, they use a flexible knife or a spring-loaded mount that absorbs the sudden resistance changes. The idea is that mechanical compliance can smooth out the force profile without needing a fast control loop. But then you lose precision — you can't do a fine dice, just coarse chops. Luna: So for a home cook who wants a robot to dice an onion into perfect brunoise, we're still years away? Lucas: I'd say at least five to seven years for a reliable consumer product. The industrial side might get there faster if we accept some waste — say, 10 percent of onions get mangled. But for a restaurant kitchen that buys onions by the crate, that's still a problem. They need consistent yield. Luna: What about using computer vision to scan the onion before cutting? Could that help? Lucas: It's been tried. Pre-scanning with an rgb d camera to estimate the internal structure. But an onion's layers aren't visible from the outside. You'd need some kind of ultrasound or near-infrared to see inside. That adds cost and complexity. And even then, the cutting process itself changes the geometry — the onion deforms as you push down. Luna: So the real challenge is real-time adaptive control with tactile feedback that can handle biological variation. Lucas: That's the nutshell. And it's not just onions. It applies to any soft, irregular food item — tomatoes, mushrooms, avocados. The food processing industry is desperate for automation that can handle this. Currently, most commercial vegetable chopping is done by high-speed blades that just smash through, creating a lot of waste. If we could solve the onion problem, we'd unlock a huge market. Luna: How big are we talking? Lucas: The global food automation market was valued at around $14 billion in 2025, and analysts project it to grow to $25 billion by 2032. A big chunk of that is in primary processing — cutting, slicing, dicing. If robots could handle onions reliably, you'd see adoption in fast-food chains, salad manufacturers, frozen food plants. Luna: And yet, a human with a knife can do it in 30 seconds with zero waste. Lucas: That's the humbling part. We take for granted how much computation is happening in our cerebellum — the proprioception, the tactile sensing, the micro-adjustments. Our fingertips have over 2,000 mechanoreceptors per square centimeter. The best tactile sensors on a robot arm have maybe 100. Luna: Speaking of interesting tech conversations — if today's episode gave you something to think about, listener support is what keeps this show ad-free and independent. If you find value in episodes like this, you can support us at buy me a coffee dot com slash fexingo. It's a small gesture that helps us keep diving into niche topics like onion chopping. Lucas: Yeah, we really appreciate the listeners who do that. It lets us spend time on deep dives instead of chasing trends. Anyway, back to the onion — one promising approach I've seen is using a combination of soft grippers and model predictive control with online learning. Instead of pre-training on a static dataset, the robot updates its model after each chop. Luna: So it learns on the job. Lucas: Exactly. A group at TU Delft published a paper last year where they used a Gaussian process regression to model the onion's stiffness in real time. They started with a prior based on generic vegetable properties, then updated the model as the knife entered the onion. After about five onions, the robot could predict the optimal force trajectory for a new onion with 90 percent accuracy. Luna: But that's still only five onions. How long until it can handle a crate of fifty? Lucas: That's the scalability question. In a lab, it works. But in a factory, you'd have many different onion varieties across seasons, different storage conditions. The model would need to adapt continuously. And if the robot makes a mistake — say, it crushes the onion — that's lost product. So there's a risk tolerance issue. Luna: It sounds like the onion is the new benchmark for dexterous manipulation, like the Rubik's cube was for multi-fingered hands. Lucas: That's a great analogy. The Rubik's cube problem was about planning and control with known geometry. The onion problem is about uncertainty and adaptation. I think we'll see more robotics labs adding 'the onion challenge' to their test suites. It's a good proxy for a whole class of real-world tasks. Luna: So what's the takeaway for someone who works in robotics or invests in automation? Lucas: I'd say: don't underestimate the difficulty of handling biological materials. The easy problems — like picking up a rigid object with known shape — have largely been solved. The next frontier is soft, deformable, variable objects. That's where the market value and the research challenges are. And the onion is a perfect test case. Luna: Alright, I'm going to go chop an onion by hand and appreciate my own fingers. Lucas: I'll join you. But I'll be thinking about the control theory while I do it.