Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Thread a Needle
Transcript
- Lucas: So here's a question: when was the last time you threaded a needle by hand? Luna: Probably a few weeks ago, fixing a button. Took me maybe ten seconds, first try. Lucas: Right. Now imagine asking a $100,000 industrial robot arm to do the same thing. It would fail. Over and over. And that failure tells us something fundamental about where robotics still is in 2026. Luna: Threading a needle — it seems so simple. What makes it so hard for a robot? Lucas: Three things. First, the hole. A standard sewing needle has an eye that's about 0.5 to 1 millimeter in diameter. Industrial robot arms have a repeatability of about 0.02 to 0.1 millimeters — so positional accuracy isn't the bottleneck. Vision systems can locate the eye just fine. Luna: So it's not about seeing the hole. Lucas: No. The second problem is the thread. It's a flexible, deformable object. When you push it, it buckles, bends, twists. It doesn't obey rigid-body physics. Robots are built for rigid objects — metal parts, plastic casings, that kind of thing. Thread is like trying to push a cooked spaghetti noodle through a keyhole. Luna: And the third problem? Lucas: Force feedback. When a human threads a needle, they feel the thread touching the metal of the eye, the slight resistance, and they micro-adjust — a fraction of a millimeter, a slight rotation of the thread. Most robot arms have force-torque sensors, but the sensitivity and speed required to do that in real time, at that scale, is still beyond what's commercially available. Luna: I remember reading a paper from MIT CSAIL last year — 2025 — where they got a robot to thread a needle with a 60 percent success rate in lab conditions. Lucas: That's the state of the art. Sixty percent. In a controlled environment, with specialized grippers, custom vision algorithms, and a lot of compute. For a factory floor that needs 99.9 percent reliability, that's not even close. Luna: And it's not just sewing. Surgical robots need to thread sutures. Textile manufacturing, shoe assembly, fishing net repair — there's a whole industry waiting for this capability. Lucas: Exactly. The economic incentive is enormous. The global textile and apparel market is worth over a trillion dollars, and a huge chunk of that is still manual labor because automation can't handle thread. Same with surgical robotics — the da Vinci system can do incredible things, but it still relies on human surgeons to manipulate sutures. Luna: So what would it take to get from 60 percent to 99.9 percent? Lucas: Better force sensing at the fingertip level — not just at the wrist. Right now, most force-torque sensors are in the robot's wrist, which gives you force at the end effector, but not at the individual finger. You need tactile sensors that can feel the thread bending at the point of contact. And you need control algorithms that can reason about flexible objects in real time. Luna: That's a machine learning problem, isn't it? Lucas: Partly. There's been good work on learning-based manipulation of deformable objects — like folding towels or tying knots — but threading a needle is a special case because the tolerances are so tight and the margin for error is almost zero. It's a precision manipulation problem with a flexible object. That's a double whammy. Luna: And it's not just about the robot arm itself. The gripper matters too. Lucas: Absolutely. Most robot grippers are designed for rigid parts — parallel jaws, suction cups, three-finger hands. For thread, you almost need a different kind of end effector. Some researchers use a small hook to pull the thread through, but that adds another layer of complexity. Luna: There's also the issue of thread variability. Different thicknesses, materials — cotton, polyester, silk. Each behaves differently. Lucas: Right. A robot trained on polyester thread might fail on silk because silk is more slippery and has different buckling properties. So you need generalization, which is still a frontier in robot learning. Luna: It's a reminder that some of the hardest problems in robotics are the ones that seem trivial to humans. Lucas: That's the theme of this whole podcast, honestly. But I think needle threading is a particularly clean example because it isolates the core challenges: precision, deformable objects, and force feedback. If we solve this, a lot of other doors open. Luna: Speaking of doors opening — and this is a bit of a left turn, but it connects — a lot of the research we're talking about comes from labs that are funded by public grants or private foundations. And one of the reasons we can cover this kind of niche, technical ground on the podcast is that we don't have to chase ad revenue. Lucas: Exactly. We keep this show ad-free, and that's only possible because of listeners who chip in a couple of bucks a month. If you've gotten something out of these episodes — maybe a new insight, a better understanding of why your robot vacuum still bumps into things — consider throwing a few dollars at buy me a coffee dot com slash fexingo. Luna: It genuinely makes a difference. Even small amounts add up and let us keep digging into these topics without having to sell you a mattress or a meal kit. Lucas: Right. So back to needle threading. One approach that's gaining traction is using visual servoing combined with a learned policy. The robot watches the thread approach the eye, and a neural network adjusts the trajectory in real time based on visual feedback. Luna: But that needs a lot of data. Thousands of attempts, I'd imagine. Lucas: Tens of thousands. And simulation can help, but the physics of thread is notoriously hard to simulate accurately. So there's a sim to real gap that's especially wide for deformable objects. Luna: What about using a different strategy altogether? Like not pushing the thread through, but pulling it? Lucas: That's exactly what some surgical robot teams are trying. Instead of feeding the thread through the needle eye, they grip the thread on the far side and pull it through. That avoids the buckling problem. But it requires a second gripper or a very precise handoff. Luna: And that adds complexity and cost. Lucas: Yeah. For a factory, doubling the grippers might be acceptable. For a surgical robot, you want the minimum number of instruments in the body. So the trade-offs are different by domain. Luna: So where do you see the first commercial application of needle-threading robots? Lucas: I think textile manufacturing will get there first. Not because it's easier, but because the cost of failure is lower. A missed stitch in a shirt is a reject; a missed stitch in a surgical suture is a life-threatening complication. So factories will adopt it sooner. Luna: And the timeline? When might we see a robot that can reliably thread a needle on a factory floor? Lucas: I'd say five to seven years for a specialized system — one that handles a specific thread type, a specific needle. General-purpose needle threading? That's probably a decade out, maybe more. Luna: That's longer than a lot of people expect. The hype around robotics always promises next year. Lucas: It does. But the physics doesn't care about hype. Threading a needle is a genuinely hard sensorimotor skill. It's a good benchmark for where we are, and a humbling one. Luna: Alright, well, next time I complain about fumbling a thread, I'll remind myself it's a frontier of robotics. Lucas: Exactly. And if you want to see a robot fail at it, there are plenty of YouTube videos. They're oddly satisfying. Luna: I'll stick to doing it myself. Thanks, Lucas. Lucas: Thanks, Luna.