Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Thread a Needle
Transcript
- Lucas: You know that feeling when you're trying to thread a needle and the thread just won't go through? You squint, you wet the tip, you try again. Now imagine a multi-million dollar industrial robot arm trying to do the same thing. Luna: I imagine it would look pretty ridiculous — but also, why would anyone want a robot to thread a needle in the first place? Lucas: That's exactly the question. The answer is the garment industry. If you could automate sewing even a fraction of the world's clothing, you'd revolutionize supply chains. But the first step is getting thread through a needle eye, and that simple action turns out to be a robotics nightmare. Luna: Okay, so what exactly makes it so hard? I mean, we have robots that can solder circuits and assemble iPhones — things that require micron precision. Lucas: Right, but those tasks involve rigid parts — a circuit board doesn't flop around. Thread is a flexible, limp, deformable object. It's a catenary with almost no stiffness. When a robot tries to push a thread into a needle eye, the thread buckles, bends, or catches on the edge. Luna: So it's a material problem as much as a control problem. Lucas: Exactly. The needle eye is typically about 500 microns wide — half a millimeter. That's roughly the width of a human hair. The thread is maybe 200 microns thick. So the clearance is tiny. And the robot has to guide the tip of that floppy thread straight through a hole that's just slightly bigger than the thread itself. Luna: And the thread isn't perfectly straight — it has a natural curl from being wound on a spool, right? Lucas: That's another layer. Thread has memory. Even if you cut it fresh, the tip will have a slight curvature. So the robot's gripper has to align the tip within a few microns of the center of the eye, at the correct angle, while moving at a reasonable speed. Luna: I've seen some factory solutions that use a vacuum to straighten the thread before inserting it. Lucas: That works for some cases, but it's not robust. A research team from MIT and the University of Washington published a paper in early 2025 showing a new approach. They used a high-speed camera running at a thousand frames per second combined with a controlled air jet to gently blow the thread straight. Then they had a needle mounted on a precision stage that moved to meet the thread, rather than moving the thread to the needle. Luna: So they flipped the problem: instead of the thread moving, the needle comes to the thread. Lucas: Yes, and that helped a lot because the needle eye is rigid. But even then, their success rate was only about 85 percent for a single thread. For multiple threads — like a double-needle sewing machine — it dropped to 60 percent. And that's in a controlled lab with perfect lighting. Luna: So not yet ready for a factory floor. Lucas: Right. And the problem gets worse with different thread types: cotton, polyester, silk — each has different friction, stiffness, and electrostatic properties. A robot that handles cotton perfectly might fail with silk because the silk is more prone to fraying. Luna: It really highlights how much we take for granted as humans. We have hundreds of thousands of years of evolution in fine motor control. Lucas: Exactly. And there's another dimension: vision. Current robot vision systems use stereo cameras or structured light to estimate pose. But thread is so thin that it barely registers on a depth sensor. So you need specialized high-resolution cameras and algorithms that can track a point that's just a few pixels wide. Luna: Plus the thread can move quickly when blown or pushed — the latency of the control loop becomes critical. Lucas: That's why the MIT team used a thousand frames per second camera. Standard industrial vision runs at 30 to 60 frames per second. By the time the robot sees the thread position, the thread has already moved. So you need a predictive model of the thread dynamics, which is an active area of research. Luna: It sounds like a fundamental problem in deformable object manipulation. Is that a hot topic in robotics right now? Lucas: Very much so. There's a whole subfield called 'robotic manipulation of deformable objects.' It includes anything from folding towels to cutting soft tissue in surgery. Threading a needle sits right at the intersection of small scale and flexibility — the hardest combination. Luna: Absolutely. In minimally invasive surgery, robots already assist by holding cameras, but suturing is still mostly done by humans. If they could reliably thread a needle and suture autonomously, it would free up surgeons' hands. Lucas: Before we dive deeper, I want to mention something that keeps our show ad-free and independent. A little support from listeners goes a long way. If you find these conversations useful, you can toss a couple of dollars our way at buy me a coffee dot com slash fexingo. It genuinely helps us keep researching topics like this and not have to run ads. Luna: Yeah, even a coffee's worth each month makes a difference — and it lets us stay focused on the stuff you care about. Lucas: Right, so back to the needle. One idea that's gaining traction is using machine learning to train robots on simulated thread dynamics. You can run millions of attempts in simulation, learn a policy that generalizes to real hardware. Luna: But simulation of flexible threads is notoriously hard — the physics is nonlinear and the contacts are complex. Lucas: Yes, but recent work from Google's Robotics division has shown that with a neural network trained on a mix of simulated and real data, a robot can achieve 70 percent success at threading a needle. That's up from 20 percent a few years ago. Luna: So the progress is real, just slow. Lucas: Exactly. And the key insight is that it's not just about better grippers or cameras — it's about understanding the physics of the thread. The thread's buckling behavior, its frictional interaction with the needle, and the way it deforms under gravity. Luna: It's almost like we need a new kind of perception — not just seeing the thread, but feeling it. Lucas: That's where haptics and force sensing come in. Some researchers are embedding tiny force sensors in the gripper to detect when the thread touches the needle's edge. You can then use that tactile feedback to align the thread. But those sensors are expensive and add complexity. Luna: Is there any simpler hack that industry uses today? Like, can they modify the needle or thread to make it easier? Lucas: Some sewing machines have automatic needle threaders that use a small wire hook to pull the thread through. But that's a mechanical solution designed for a specific needle geometry. For a general-purpose robot arm, you want to handle any needle and any thread. Luna: So the real challenge is generalization. A robot that works for one setup might fail when the thread or needle changes. Lucas: That's the frontier. And it's why this problem is so fascinating. It's not a hardware limitation per se — it's a lack of fundamental understanding of how to manipulate flexible materials. Luna: Do you think we'll see a robot that can reliably thread a needle in the next five years? Lucas: If you ask me, probably in specialized applications — like a robot that only uses one type of thread and needle — yes, I think we'll get there. For a general-purpose robot that can pick up any needle and thread from a tray, that's a decade out at best. Luna: So for now, humans still have this one over the machines. I'm okay with that. Lucas: Me too. It reminds us that our own fine motor skills are an incredible piece of biological engineering.