Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / How Robot Arms Learned to Cut a Lemon
Transcript
- Lucas: So — robot arms can weld a car chassis to within a tenth of a millimeter, but ask one to cut a lemon in half and suddenly the thing looks like it's trying to perform surgery with a sledgehammer. Luna: I've seen the videos. The arm either crushes the lemon into pulp or barely scratches the peel. Why is cutting so much harder than welding? Lucas: Welding is rigid — the parts are fixed, the torch trajectory is pre-programmed, the material doesn't deform. A lemon? It's soft, it's slippery, it has a variable internal structure. The blade has to find the exact boundary between peel and flesh, and that boundary changes with every single fruit. Luna: So it's not just about vision? I assumed a camera could locate the center and the arm could just slice down. Lucas: Vision gets you part of the way — you can estimate the overall shape, maybe the stem axis. But the lemon's interior isn't visible. The pith thickness varies, the seeds shift, the fruit might be slightly overripe or underripe. A static cut depth based on a camera image will fail on a non-standard lemon. You need real-time force feedback — the arm has to feel its way through the cut. Luna: So it's a tactile intelligence problem, not a vision problem. What does the force feedback actually look like in practice? Lucas: A typical six-axis industrial arm, say a Universal Robots UR10, has joint torque sensors that measure resistance at each joint. When the blade contacts the lemon peel, the resistance spikes. A naive controller tries to maintain a constant downward force — but that leads to crushing because the lemon's stiffness changes non-linearly as the blade cuts through. The breakthrough was using impedance control: the arm modulates its stiffness and damping in real time based on the measured force and position. Luna: Impedance control — that's the idea that the arm behaves like a spring-damper system, right? It can be 'soft' when it needs to adapt, and 'stiff' when it needs to cut through? Lucas: Exactly. A group at MIT CSAIL published a paper last year where they trained a reinforcement learning policy on a simulated lemon model with randomized pith thickness and ripeness. The policy learned to start with a high stiffness for the initial puncture, then relax stiffness as it approached the inner flesh, and finally stiffen again when it hit the cutting board. They transferred the policy to a real Mitsubishi RV-12 robot with a custom force-torque sensor at the wrist. Success rate on lemons? 94 percent. That's up from about 40 percent with a fixed-force baseline. Luna: 94 percent is impressive. But is it generalizable? Would the same policy work on a tomato or an avocado? Lucas: That's the open question. The policy was trained specifically on lemons, so the force signatures it learned are tuned to citrus. A tomato has a much thinner skin and a softer interior — the same policy would likely over-penetrate and cause crushing. But the researchers showed that the simulation to reality transfer framework itself is general. You can swap in a new fruit model, retrain for a few thousand episodes in simulation, and deploy. The bottleneck becomes the simulation fidelity — you need an accurate FEM model of the fruit's material properties. Luna: That sounds computationally expensive. Is there a simpler approach that doesn't require full finite element modeling? Lucas: Some groups are using tactile sensors on the gripper — essentially an artificial skin that maps pressure distribution across the contact area. GelSight sensors, for example, use a camera and a deformable gel to reconstruct surface geometry and force at high resolution. With those, you can estimate the fruit's firmness from the initial contact and adapt the cut parameters without a full simulation. It's less precise than fem based RL, but more practical for real-world deployment. Luna: But even with tactile sensing, the arm still needs to know when to stop. How does it detect that the cut is complete? Does it sense the change in force when the blade hits the cutting board? Lucas: That's actually the hardest part. The board provides a sharp force discontinuity — but by then the blade might have already over-traveled and damaged the fruit. The MIT policy learned to anticipate the board contact by detecting a characteristic force profile about 2 millimeters before full separation. It then reduces velocity and force to minimize crushing on the exit. It's a predictive model, not just reactive. Luna: That's remarkably human-like. A person doesn't think about it — they just feel when the knife is through. But replicating that 'just feeling' is incredibly complex. Lucas: It is. And this is why the lemon-cutting problem has become a benchmark in the manipulation community, similar to how the peg-in-hole task was the classic test for assembly. It forces the arm to handle non-rigid contact, variable geometry, and real-time adaptation — all at once. Luna: So if you can cut a lemon reliably, you can probably cut many other deformable objects. Where does this technology go next? Commercial food processing? Lucas: Yes, but also surgical robotics. The same impedance control framework applies to cutting soft tissue — a surgeon's scalpel needs to adapt to different tissue densities. In fact, one of the MIT co-authors is now working with a medical robotics startup on a needle insertion system that uses similar force-feedback policies to avoid puncturing blood vessels. Luna: That's a direct transfer. From lemon to liver. Lucas: Exactly. And the agricultural applications are huge. Lettuce harvesting, for instance — you need to cut the stem at the right height without damaging the leaves. Current robotic harvesters use vision alone and have a high failure rate. Force feedback could dramatically improve yield. Luna: There's also the packaging industry. I've seen videos of robots trying to pick up a single potato chip from a conveyor belt without breaking it. That's another deformable object problem. Lucas: Chip handling is a nightmare because the geometry is random and the chips are brittle. But force feedback can help there too — some research uses vacuum grippers with airflow sensors to detect when the chip is fully sealed, then adjust lift speed based on chip weight. Luna: So we're really talking about a fundamental shift in how robot arms interact with the physical world. Moving from position control — 'move to this exact coordinate' — to force control — 'apply this much force while moving'. Lucas: Right. And that shift is happening now because sensors and compute are cheap enough. A six-axis force-torque sensor used to cost $10,000. Now you can get a decent one for under $2,000. Combined with off-the-shelf collaborative robots, small and medium manufacturers can start experimenting with force-controlled tasks. Luna: That makes the technology more accessible. But I imagine the software stack is still the hard part. Writing impedance control laws isn't trivial. Lucas: It's getting easier. ROS, the Robot Operating System, has packages for force control that abstract away a lot of the low-level math. And companies like Franka Emika have built impedance control directly into their robot's low-level controller — you can set stiffness and damping parameters in real time via a simple API. Luna: If today's episode gave you a useful window into the tactile challenges in robotics, I'll just mention that these episodes stay ad-free thanks to listener support. If you find value in deep dives like this, you can buy me a coffee at buy me a coffee dot com slash fexingo. It's a small gesture that keeps the show independent. Lucas: And it genuinely helps us spend more time on topics like force control and impedance learning — which don't always get mainstream coverage. So if you've been enjoying the show, consider it. And now back to the lemon. Luna: So we've got a robot that can cut lemons with 94 percent accuracy. Is there any commercial deployment of this yet, or is it still lab-scale? Lucas: It's mostly lab-scale, but there are pilot projects. A Japanese food processing company — let me think — Nippon Ham, I believe, is testing a force-controlled cutting system for boneless ham slicing. The goal is to adapt slice thickness based on meat density variations, which improves yield by about 5 percent. That's significant at industrial volumes. Luna: Five percent yield improvement on a product line cutting millions of slices a day — that's a lot of profit. Lucas: Exactly. So the business case is already there for high-value food products. The challenge is making the system robust enough to run 24/7 without an engineer tuning parameters. The lemon-cutting policy works 94 percent of the time in a lab, but 94 percent is not enough for a factory floor — you'd need 99.9 percent plus graceful failure recovery. Luna: What does failure look like? Does the arm just crush the fruit, or can it stop and retry? Lucas: In the current system, failure typically means either under-cutting — the blade stops too early and you have a partially cut lemon — or over-cutting, where the blade goes through the board. Neither is catastrophic, but they waste product. The researchers are working on a failure detection classifier that runs concurrently with the cutting policy — if the force profile diverges from expected, the arm backs out and repositions for a second attempt. Luna: That adds a lot of complexity. But it's the kind of robustness you need for real-world use. Lucas: It's the difference between a demo and a product. And that gap is where most robotics research stalls. But the fact that we're now having serious conversations about cutting lemons — a task that seemed impossible five years ago — tells you how fast the field is moving. Luna: What's the next fruit on the benchmark list? I've heard rumors of a 'banana peeling' challenge. Lucas: Banana peeling is actually active research at UC Berkeley. The problem there is that the peel has a natural seam — you need to detect the seam and then apply a peeling motion that separates the peel without breaking the fruit. It's a multi-step manipulation task combining vision, force, and fine motor skills. Very much an open problem. Luna: So we've gone from 'robot can't handle a single grape' to 'robot can cut a lemon'. That's real progress. Lucas: It is. And the key takeaway for me is that the bottleneck isn't hardware — it's software. The arms themselves have had the necessary precision and torque bandwidth for years. What was missing was the control methodology to handle uncertainty. Reinforcement learning, combined with simulation, is finally providing that. Luna: It's a good reminder that the hardest problems in robotics aren't about strength or speed — they're about touch. Lucas: Exactly. And as force sensors get cheaper and simulation models get more accurate, I think we'll see robots handling deformable objects as naturally as they handle rigid ones. Maybe in five years, that fitted sheet folding problem won't seem so impossible. Luna: That's a hopeful note to end on. Thanks, Lucas. Lucas: Thanks, Luna. That's it for this episode of The Robotics Podcast.