Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Pick Up a Single Piece of Spaghetti
Transcript
- Lucas: So, we've spent dozens of episodes watching robot arms fail at things that seem simple to us — eggs, towels, zippers. Today I want to talk about one that really broke my brain when I first saw the research. Luna: Okay, I'm ready. What's the new failure? Lucas: Picking up a single piece of dry spaghetti. Not a handful. Not a bundle. One strand. Luna: Wait — like, from a flat surface? Just a single straight piece? Lucas: Exactly. A typical piece of dry spaghetti is about 2.5 millimeters thick and 25 centimeters long. It's brittle, it's slightly curved, and it has a very specific ratio of length to diameter. For a robot, that's a nightmare. Luna: I can see how it's fragile, but is it really that much harder than a screw or a lightbulb? Lucas: It's a different kind of hard. With a screw, you can align the gripper perpendicular to the axis and apply a controlled force. But a spaghetti strand isn't rigid — it's what roboticists call a 'slender deformable object.' When you try to grab it near the center, it bends. When you try near the end, it might slide away. And if you apply too much force, it snaps. Luna: So the robot has to find exactly the right grip point and apply exactly the right pressure, all while dealing with the fact that the object itself is slightly curved and might roll. Lucas: Right. In a 2024 paper from MIT's Computer Science and Artificial Intelligence Lab, researchers tested a standard industrial robot arm with a parallel-jaw gripper on 50 different strands. The success rate was 12 percent. Human volunteers got 94 percent. Luna: Twelve percent. That's embarrassingly low. What went wrong most often? Lucas: Three failure modes. First, the strand would roll away from the gripper — because the surface is smooth and the contact area is tiny. Second, the gripper would pinch too hard and snap the pasta. Third, the robot would try to pick it up, but the strand would just slip out because the friction was too low. Luna: So it's a combination of sensing, control, and mechanical design. This feels like a fundamental problem in manipulation. Lucas: It is. And it gets at something deeper: human hands have thousands of tactile sensors, dynamic stiffness modulation, and years of practice. A robot arm has maybe a force-torque sensor at the wrist and a camera. It's blind to the subtle feedback you get from your fingertips. Luna: Have any approaches worked better? Like, what about using a vacuum gripper? Lucas: Vacuum grippers do better on flat, smooth surfaces. But a spaghetti strand is narrow and curved — you'd need a tiny suction cup and perfect alignment. In that same study, they tried a soft gripper with three fingers, and success rate went up to 38 percent. Luna: Better, but still far from reliable. So what's the research community learning from this? Lucas: One big lesson is that you can't just brute-force it with better vision or stronger motors. You need what roboticists call 'compliant manipulation' — the ability to adjust grip force in real-time based on tactile feedback. There's a group at Stanford working on a gripper that uses piezoelectric sensors to measure micro-slips before the object actually moves. Luna: That's interesting. Because if you can detect that the strand is about to slip, you can increase pressure just enough to stop it, but not so much that you snap it. Lucas: Exactly. The sensor detects the slip onset — that's a few milliseconds before the object actually escapes. The controller then adjusts the grip force continuously. It's similar to how your own fingers work: you don't squeeze a glass with constant force; you modulate it based on how heavy it feels and how much it's moving. Luna: So the spaghetti problem is essentially a test bed for advanced tactile sensing and control. I imagine the applications go way beyond pasta. Lucas: Absolutely. If you can reliably pick up a single strand of spaghetti, you can pick up a wire, a fiber-optic cable, a suture thread, or a delicate electronic component. These are real challenges in manufacturing, surgery, and electronics assembly. Luna: It's funny — we think of robots as being incredibly precise, but they fail at tasks that a three-year-old can do. And that gap tells us a lot about what's still missing. Lucas: Yeah, and I think this is one of those episodes where listeners might wonder why we spend time on such a niche problem. But the truth is, every time we solve one of these 'simple' failures, we push the entire field forward. Luna: Speaking of pushing forward — I know we keep these shows ad-free by design. And a couple of dollars a month is genuinely what keeps these going. If today's tech conversation gave you something usable, you could head to buy me a coffee dot com slash fexingo. Lucas: Yeah, it's a small thing that makes a real difference for us. No pressure at all, but if you value what we do here, that's the simplest way to keep it running. Luna: Alright — back to the spaghetti. So where are we now, practically? Is this still a lab curiosity, or are there commercial systems that can handle it? Lucas: It's mostly still in the lab. There's a startup out of Carnegie Mellon called DexArt that's working on a gripper combining suction and parallel jaws. They claim 70 percent success on spaghetti in their latest preprint. But that's on a controlled surface with optimal lighting. In a factory environment? Probably much lower. Luna: So what's the next big hurdle once you solve the single-strand pickup? Bundles? Spaghetti that's stuck together? Lucas: Exactly. After single-strand, the next challenge is separating one strand from a pile without disturbing the others. Then you have to handle cooked spaghetti, which is sticky and even more deformable. Each level adds complexity. Luna: I'm guessing the cooked version is way harder because the surface is slippery and it sticks to itself. Lucas: Right. Cooked spaghetti has a different coefficient of friction, it's compliant, and it can entangle. There's a 2025 paper from the University of Tokyo that tried to pick up cooked strands using a vision-based reinforcement learning approach. They had a 3D printer custom-make a gripper with soft fingers. Success rate: 22 percent. Luna: So even with a custom gripper and machine learning, still only one in five. This really highlights how far we are from human-level dexterity. Lucas: It does. And it's humbling because the human hand is an incredible piece of engineering. We have about 20 muscles in the hand alone, plus thousands of mechanoreceptors. A typical industrial gripper has one or two degrees of freedom and maybe a dozen sensors. Luna: So what's the current state of the art? Is there any robot that can reliably pick up a single strand of dry spaghetti in a real-world setting? Lucas: As of mid-2026, not really. The best systems are around 70 to 80 percent in lab conditions, but they require precise calibration, controlled lighting, and often multiple attempts. For practical applications, you'd need 99.9 percent reliability. Luna: That seems a long way off. But I suppose every failure teaches you something about the fundamental physics of manipulation. Lucas: Exactly. And that's why I find this problem so fascinating. It's not about pasta. It's about understanding how to interact with a world full of slender, fragile, deformable objects. The spaghetti is just a perfect test case.