Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Thread a Needle
Transcript
- Lucas: You know what I did over the weekend, Luna? I darned a sock. One of my good wool socks had a hole near the heel, so I threaded a needle, made a few stitches, and fixed it. Took maybe two minutes. Luna: Okay — I'm impressed. But I suspect this is going somewhere. Lucas: It is. Because while I was sitting there with the needle and thread, I started thinking: this is the kind of task we assume is trivial. And then I remembered that we've done over a hundred episodes on what robot arms can't do, and we've never covered threading a needle. Luna: Huh. That's a good point. I can tie a knot in under a second, but we just did an entire episode on why robot arms can't tie knots. Lucas: Exactly. So today, let's talk about why threading a needle is genuinely hard for a robot. And I don't mean 'hard' in the sense of 'we need better code.' I mean there are fundamental physical and sensory challenges that make it a frontier problem in dexterous manipulation. Luna: Alright. So what's the first thing that makes it so tough? Lucas: Let's start with the geometry. A standard sewing needle has an eye that's maybe half a millimeter wide. The thread is about 0.3 millimeters in diameter. So the clearance — the gap between the thread and the hole — is roughly 0.1 millimeter all around. That's the width of a human hair. Luna: Okay, so it's a precision alignment problem. But industrial robot arms have repeatability down to like ten microns. So why can't they just align it? Lucas: You'd think so. But here's the catch: the robot doesn't know where the needle eye is in real time. It might have a camera, but the camera updates at 30 frames per second. And the robot's controller has latency from image processing, path planning, and actuator response. By the time the robot moves to where the eye was, the eye — if the needle moves even slightly — is somewhere else. Luna: So it's a timing issue. The robot is trying to hit a moving target. Lucas: Right. And the human solution is tactile. When I thread a needle, I don't just use my eyes. I feel the thread contacting the metal. I feel the slight resistance when the tip of the thread hits the edge of the eye. I adjust based on that tactile feedback — micro-movements, maybe a couple hundred microns at a time. Most robot arms don't have that sense of touch. Luna: There are force-torque sensors at the wrist. Couldn't those pick up the contact? Lucas: They can, but the forces involved in threading a needle are incredibly small. The thread is flexible and lightweight. The contact force when the thread brushes against the needle might be in the order of millinewtons. A typical industrial force-torque sensor has a resolution of about 0.1 newtons. So it's an order of magnitude too coarse to feel that interaction. Luna: So you'd need a much more sensitive sensor. And maybe closer to the end effector, not at the wrist. Lucas: Exactly. That's why there's been a lot of work on tactile sensors — like the GelSight sensor developed at MIT. It's essentially a gel-filled pad with a camera that images deformations at high resolution. It can detect forces in the millinewton range and even feel texture. But integrating that into a robot gripper is still research-stage. Luna: And even then, you've got the flexibility of the thread. A thread isn't a rigid object. It buckles, bends, and twists. How do you control something that's effectively a string? Lucas: That's the other half of the problem. Manipulating a thread is a classic example of 'dexterous manipulation with flexible objects.' The robot has to not only align the tip but also keep the thread from buckling. Humans do this by holding the thread close to the tip — maybe a centimeter away — and maintaining tension. But that requires coordinating two hands: one holding the needle, one holding the thread. Most industrial setups are single-arm. Luna: Could you do it with two robot arms? Like a dual-arm collaborative robot? Lucas: Theoretically, yes. And there's been research on that. In 2025, researchers at MIT's CSAIL published a paper specifically on robot needle threading using two arms: one with a needle holder, one with a thread gripper. They used a combination of high-speed cameras at 200 frames per second and a force-sensitive gripper on the thread side. Their success rate was about 60 percent for a single threading attempt, compared to basically 100 percent for a human. Luna: Sixty percent — that's not bad. But it's not production-ready. Lucas: Not for a task where failure costs time. And the cycle time was about 12 seconds per attempt. Humans do it in two or three seconds. So it's not just about success rate; it's about speed. And the robot struggled when the thread was slightly twisted or when the lighting changed. So it's far from robust. Luna: Okay, so let's zoom out. Why does this matter beyond darning socks? Who actually needs robots to thread needles? Lucas: Several high-value industries. One is surgical robotics. In minimally invasive surgery, you often need to suture — stitch tissue together — inside the body, through tiny incisions. The needle is curved and the thread is attached. Having a robot that can autonomously thread a needle through a suture driver would be a big step toward automating parts of surgery. Luna: Right, because the da Vinci system still relies on the surgeon to do the threading manually, right? Lucas: Yes. The da Vinci system replicates the surgeon's hand movements, but the surgeon is still controlling every step. If the robot could autonomously thread a needle, it could potentially perform sutures without direct human control — which could be useful in tele-surgery or in scenarios where bandwidth is limited. Luna: Interesting. What else? Lucas: Another big application is assembling fiber-optic connectors. Those connectors have tiny holes — sometimes 125 microns in diameter — that need to align with a fiber. It's essentially the same problem: threading a flexible strand into a small hole. Currently, that's done by skilled technicians under microscopes. If a robot could do it reliably, it could speed up manufacturing and reduce costs. Luna: So the needle-threading problem is really a micro-assembly problem. It's just a convenient benchmark. Lucas: Exactly. And that's why it's been studied for decades. There's a famous 1996 paper by a researcher named H. Van Brussel at KU Leuven on 'assembly of flexible parts' that used needle threading as a test case. But the core challenge hasn't changed: you need high-speed visual feedback, sensitive tactile sensing, and fine force control — all at once. Luna: Has anyone tried using event-based cameras instead of traditional frame-based cameras? Lucas: Yes. And that's actually a promising direction. Event-based cameras detect changes in brightness per pixel, asynchronously, at microsecond resolution. They don't have the latency of a 30-fps camera. So if the needle moves, the camera registers it almost instantly. Some recent work from the University of Zurich showed that event-based cameras can reduce the alignment error in a needle-threading task by about 40 percent compared to a standard camera. Luna: Forty percent is significant. But you still need the tactile feedback, right? Because vision alone can't tell you if the thread is buckling. Lucas: Right. And there's also the question of the gripper. Most robotic grippers are designed for rigid objects. Holding a needle is one thing — you can use a vice-like grip. But holding a thread gently enough not to crush it, but firmly enough to control it, is tough. Some groups are using soft grippers with pneumatic actuation, but they're slower. Luna: So what's the state of the art today? In 2026, has anyone actually built a robot that can thread a needle reliably? Lucas: Not in a robust, commercial sense. The best lab results I've seen — from a 2026 preprint out of the German Aerospace Center — achieved about 85 percent success rate with a custom two-fingered gripper equipped with tactile sensors and a high-speed camera running at 500 Hz. But they needed carefully controlled lighting, the thread had to be pre-cut to exactly 5 centimeters, and the needle had to be fixed in a holder. It's still a lab demo. Luna: So it's like catching a ball or folding a fitted sheet — we're getting closer, but it's not solved. Lucas: Exactly. And that's what makes this such a good benchmark for robotic dexterity. It combines so many hard problems: visual servoing, force control, flexible object manipulation, and bimanual coordination. Every time someone makes progress on one of those fronts, the needle-threading success rate goes up a few percentage points. Luna: I wonder if, in ten years, we'll look back and think 'remember when robots couldn't thread a needle?' the way we now think about chess or Go. Lucas: I think so. And it might happen sooner than we expect, because the economic incentives are there. If you can automate fiber-optic assembly, that's a huge market. So there's real money pushing this forward. Luna: Alright, before we wrap up — one thing I'm curious about: you mentioned the human tactile feedback. Could we give a robot a sense of touch that's actually better than human? Like, could we put sensors on the thread itself? Lucas: That's a fascinating idea. Some researchers have experimented with 'smart thread' — thread coated with conductive material so that when it contacts the needle, you get an electrical signal. That gives you binary contact detection, but not force magnitude. There's also work on using acoustic sensors to listen for the sound of thread rubbing against metal. But it's all experimental. Luna: So we're not there yet. But the podcast community is — and that's a nice segue. You know, we try to keep Fexingo Business ad-free — no sponsors, no interruption. And that's only possible because of listeners who support us directly. If today's episode on robot dexterity gave you a new way to think about automation, and you'd like to help us keep these conversations going, you can find us at buy me a coffee dot com slash fexingo. It's a simple way to chip in, no strings attached. And it really does make a difference in keeping the show independent. Lucas: Yeah, we don't do this for the money — but the support from listeners is what lets us spend the time digging into topics like why threading a needle is so hard. So if you've enjoyed the show, consider it. And now back to the thread — or the needle, I guess. Luna: So, to tie it all together — no pun intended — needle threading is a perfect example of a task that looks simple but exposes the limits of current robotics. And every time we solve one piece, we get a little closer to robots that can do the kind of fine manipulation we take for granted. Lucas: And maybe one day, I'll be able to tell my robot to darn my socks. But until then, I'll keep doing it myself. Luna: I'll stick to buying new socks. Lucas: That's the human advantage — knowing when to outsource.