Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Tie a Knot
Transcript
- Lucas: So this is going to sound like a trick question, but I promise it's not: when was the last time you tied a knot? Luna: Tied a knot? This morning, actually. Shoe laces. Double knot. Lucas: Exactly. You did it without thinking. A few seconds of finger motions, and you've got a secure knot. Now ask a robot arm to do the same thing, and it falls apart. That's our topic today — why tying a knot is still a frontier problem in robotics. Luna: Okay, but we've done a ton of episodes on robot manipulation — folding shirts, screwing in lightbulbs, threading a needle. How is knot tying different? Lucas: Great question. The short answer: a knot isn't a rigid object. It's a flexible, continuous string that changes shape as you manipulate it. Most robot grippers are designed to pick up solid objects — a block, a screwdriver, a coffee cup. But a piece of string? It has no fixed geometry. It deforms, it slips, it tangles. And a knot requires you to manage multiple loops simultaneously. Luna: So the robot can't just 'grasp' the string the way it would a tool. Lucas: Right. And it's worse than that. To tie a knot, you need to create a loop, pass the end through, and then tighten — all while maintaining tension on the right segments. That's a sequence of continuous contact events. If at any point the string goes slack, the knot unravels. If the gripper pinches too hard, the string can't slide. It's a delicate dance between grip force and friction. Luna: I'm picturing a robot trying to tie its shoelaces and just yanking the whole thing into a tangled mess. Lucas: That's very close to what happens. In research labs, they often start with a thicker cord — like a shoelace or a rope — because it's easier to track visually. But even then, the failure rate is high. A 2022 paper from MIT's Computer Science and Artificial Intelligence Lab — CSAIL — reported that their system could tie a simple overhand knot about 50 percent of the time. Fifty percent. For a five-year-old human, it's basically 100 percent. Luna: What's the state of the art today, in 2026? Have we improved that number? Lucas: We have. Researchers at UC Berkeley's AUTOLab published results earlier this year showing a robot that can tie an overhand knot with about 85 percent success. But here's the kicker — it takes the robot around 45 seconds to do it. And it requires a specialized gripper with tactile sensors on every finger pad. Luna: So the robot is slower and less reliable than a human, and it needs custom hardware. Where's the practical application? Lucas: That's the key question. The most immediate use case is surgical robotics. In the operating room, surgeons tie knots with sutures all the time — it's a fundamental skill. But if a robot can tie a knot inside the body, through a tiny incision, that's a huge leap for minimally invasive surgery. There's also interest from the fishing industry — automated systems that tie fishing line to hooks and lures at scale. Luna: And I imagine each of those applications has different constraints. Surgical sutures are thin and slippery; fishing line is stiff and prone to kinking. Lucas: Exactly. One size does not fit all. And this brings up a broader point: the robotics community is realizing that 'manipulation' isn't one skill — it's a whole taxonomy. Grasping a rigid object is one thing. Handling cloth, paper, string — each requires its own physics model. Knot tying sits at the intersection of deformation, friction, and sequential actions. Luna: So what's the breakthrough that got Berkeley to 85 percent? Was it better hardware or better software? Lucas: Mostly software. They used a technique called 'deep reinforcement learning with sim to real transfer.' They trained the robot in simulation — thousands of virtual knot-tying attempts — and then transferred that policy to the real robot. The simulation allowed the system to fail millions of times without damaging hardware. And they added tactile feedback so the robot could 'feel' whether the string was sliding. Luna: That tactile part seems crucial. If you can't see exactly where the string is — because it's occluded by the robot's own hand — you need to sense it. Lucas: Exactly. Vision alone isn't enough. In my conversation with a researcher from the lab, she told me that the robot would sometimes have the knot visually correct but the loop was actually twisted — and it would pull apart. The tactile sensors caught that. So the next frontier is integrating multiple senses — vision, touch, and even force sensing at the wrist. Luna: I want to pivot to something you said earlier about the surgical use case. Is there a risk that a robot tying a suture inside a patient could fail — 15 percent of the time, according to Berkeley's numbers? That's not acceptable in the OR. Lucas: You're absolutely right. Surgical robots today — like the da Vinci system — are teleoperated. The surgeon controls every move. The knot-tying research is aimed at autonomy, but we're years away from a robot tying a suture unsupervised. The goal is more like: the robot assists by holding tension on one end while the surgeon ties, or the robot ties a simple knot that the surgeon then checks and tightens. Think of it as a collaborative tool, not a replacement. Luna: That makes more sense. So the robot handles the tedious part, and the surgeon does the quality check. Lucas: Precisely. And there's another interesting angle: in industrial settings, robots already tie knots in certain applications — like binding cables in automotive wiring harnesses. But those are highly structured environments where the wire is stiff and the knot pattern is fixed. The general problem — tying an arbitrary knot with any type of string — is still unsolved. Luna: So we have a robot that can tie an overhand knot 85 percent of the time in a lab. What about a bowline? Or a figure-eight? Or a surgeon's knot? Lucas: Each additional complexity multiplies the difficulty. The bowline, for example, requires forming a loop, passing the end up through the loop, around the standing part, and back down. That's four distinct steps with precise timing. As of mid-2026, no published paper has demonstrated a robot tying a bowline autonomously with more than 30 percent success. The figure-eight is slightly easier — around 60 percent — because it's a more symmetrical knot. Luna: And a surgeon's knot — which is essentially an overhand with an extra twist — is probably harder because it requires more wraps. Lucas: Exactly. More wraps means more friction, more chances for the string to slip. But there's some promising work from a team at the University of Washington that uses a dual-arm setup — one arm holds the suture, the other ties the knot. They reported 90 percent success on a surgeon's knot with thick suture material. Thin sutures — the kind used in eye surgery — are still below 50 percent. Luna: It's fascinating that something so simple for humans remains so hard for robots. It really highlights how much our brains and hands are optimized for this kind of task. Lucas: Absolutely. And this is why I think knot tying is such a great benchmark for robotic dexterity. It's a task that requires planning, sensing, and fine motor control all at once. If we can solve it, we'll have a much better understanding of how to build robots that can work alongside humans in messy, unstructured environments. Luna: Before we wrap up, Lucas — I know we don't run ads, and I appreciate that our listeners value that. If today's conversation gave you something useful, and you want to support keeping this podcast independent and ad-free, you can find our page at buy me a coffee dot com slash fexingo. No pressure, just a way to help us keep going. Lucas: Yeah, I'll second that. We've had a lot of listeners ask how they can support the show, and that's the simplest way. It really helps us focus on the content rather than chasing sponsors. So thank you to anyone who's contributed. Luna: Alright, back to knots. Lucas, you mentioned that the Berkeley system took 45 seconds. Is there any work on making it faster? Lucas: Yes, there is. A group at Stanford published a preprint last month showing a robot that can tie an overhand knot in under 10 seconds — but only with a specific type of stiff rope. They used high-speed cameras and a custom gripper that moves the string like a pair of hands. The trade-off is that it's less adaptable. Change the rope thickness, and it fails. Luna: So speed vs. generality. That seems like the classic robotics trade-off. Lucas: It really is. And I suspect the next few years will bring systems that combine fast, specialized motions for known knot types with slower, more sensor-driven approaches for novel ones. We're not there yet, but the pace of progress is encouraging. Luna: Where do you see this going in the next five years? Lucas: I think we'll see the first commercial surgical robot that can tie a knot autonomously — for a specific type of suture — within three years. The fishing industry automation might come even sooner, because the cost of failure is lower. And in the longer term, I expect knot-tying capability to become a standard feature of general-purpose robot hands, like the ones being developed by companies like Shadow Robot or the DLR. Luna: It's one of those skills that seems trivial until you try to program it. I'm glad we dug into it. Lucas: Me too. And if any listeners are working on this problem, we'd love to hear from you. Next week, we're talking about another deceptively hard task: why robots still can't fold a fitted sheet. Luna: Oh, that one's personal. I'll have some strong opinions. Lucas: I'm counting on it. Until next time.