Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Tie a Knot
Transcript
- Lucas: So, Luna, I've got a new one for the 'why can't robots do this' file. Today: tying a knot. Luna: Oh, a classic. But haven't we done a bunch of those? Like peeling a potato, cracking an egg, threading a needle? Lucas: Right, but those are all about individual manipulation challenges. A knot is something else entirely. It's continuous contact, two flexible strands, and the outcome depends on the sequence of moves and the tension applied at every step. It's almost like a dance. Luna: And it's not just one kind of knot either. There's the square knot, the surgeon's knot, the bowline. Each has its own geometry and required forces. Lucas: Exactly. And that's why it's such a beautiful test case for robotic manipulation. It forces you to solve perception, planning, and control all at once. A lot of the classic problems we've talked about on this show — picking a screw from a bin, folding a towel — they're hard, but they have a defined endpoint. With a knot, the endpoint is the knot itself, but getting there requires you to manage a flexible object through a series of loops and pulls. Luna: And if you pull too hard at the wrong moment, you just make a mess. A bird's nest, as sailors call it. Lucas: Right. So let's actually anchor this in a real application, because the most compelling reason to care about knot-tying robots is surgery. In minimally invasive surgery, surgeons need to suture and tie knots with instruments through small incisions. It's a skill that takes years to master, and it's physically demanding. If a robot could do it reliably, it could free up the surgeon's attention for higher-level decisions. Luna: And there's a huge incentive for that. The global market for surgical robots is already tens of billions of dollars, and it's growing double digits every year. So it's not just an academic exercise. Lucas: No, it's not. But here's the thing: even with all that money, knot-tying remains one of the hardest tasks for surgical robots. Why is that? Let's break it down. First, perception. The robot needs to know exactly where the suture thread is in 3D space, and it's a thin, flexible, often wet line. Standard cameras struggle with that. You need specialized vision that can track the thread even when it's partially occluded by the instruments or tissue. Luna: And then there's the contact problem. The thread isn't rigid, so the forces you apply at one point deform it along its entire length. It's not like picking up a block. Lucas: Precisely. That's what makes it a 'continuous contact' problem, to use the jargon. You can't just plan a trajectory and execute it open-loop. You need feedback — tactile feedback — to know how much tension is on the thread, whether it's about to slip, whether the loop is tight enough. Luna: And that's what makes it so different from, say, a robotic arm on an assembly line, where the parts are rigid and the environment is controlled. Lucas: Exactly. And that's why the recent breakthroughs have come from combining reinforcement learning with tactile sensors. There was a paper earlier this year — I think it was from a university in Japan — where they used a dual-arm robot to tie a square knot in a rope, and they did it entirely with learned policies. The robot wasn't programmed with a step-by-step instruction; it learned through trial and error, using tactile feedback to adjust its grip and tension. Luna: That's impressive. But how close is that to clinical reality? Lucas: Still a gap. The robot in that study was using a thick cotton rope, not a monofilament suture thread that's a fraction of a millimeter thick. And it was in a simplified simulation-like environment, not a real surgical site with tissue and fluids. So it's a proof of concept, not a product. Luna: But it's a start. And we've seen this pattern before: research lab success, then a long grind to make it robust and safe for real-world use. Lucas: Yeah, that's the typical robotics curve. And it's interesting to compare knot-tying to something like the warehouse pick and place problem, which has largely been solved at scale. The difference is that pick and place involves rigid objects with known geometry, and the environment is structured. Knots are the opposite: unstructured, deformable, and highly sensitive to initial conditions. Luna: And that's why we keep coming back to these 'why can't robots do X' episodes. It's not that they're impossible; it's that they require a fundamentally different approach than what works in factories. Lucas: Right. And that's what makes this so fascinating. Let's talk about another application: marine robotics. Underwater, knots are used in fishing nets, mooring lines, and scientific equipment. A robot that could tie a secure knot underwater would be incredibly valuable — for autonomous repair, for deploying sensors, for cleaning up debris. Luna: But underwater, you have water currents, limited visibility, and the robot's own movements disturb the water. That's even harder than surgery, if you can believe it. Lucas: Definitely. And there's also the completely mundane but surprisingly tricky problem of tying shoelaces. I saw a video of a robot that can do it — one of the research labs, I think it was in Italy — and it takes about ten seconds for a single shoe, and it's a very deliberate motion. It's not something you'd want to wait for if you're late for work. Luna: But it's a great demonstration. And I think that's part of the appeal of these challenges: they're relatable. Everyone has struggled with a stubborn knot, so we can intuitively appreciate why it's hard. Lucas: Absolutely. And speaking of things that are hard, I want to take a quick step back, because if today's conversation has given you something useful — and I really think it has, because we're covering a fundamental challenge in robotics that's going to unlock a lot of future applications — the way we keep these episodes free and accessible is through listener support. If you found value in this, and you want to help us keep going, you can buy us a coffee at buy me a coffee dot com slash fexingo. Luna: Yeah, every little bit helps. And it genuinely keeps us ad-free, so we can dive deep into topics like this without worrying about sponsors. Lucas: Exactly. And we're not asking for a lot — just if you've gotten something out of the show, a coffee is a great way to say thanks. Anyway, back to the knots. So we've talked about perception and control, but there's another layer: the physics of the knot itself. Luna: You mean like the topology? How a knot is actually a mathematical structure? Lucas: Yes, and that matters more than you might think. A knot is defined by the crossing pattern of the strands, and to tie it, you have to create those crossings in a specific order. It's not just about moving the thread; it's about creating the right geometry. That's a planning problem that's quite different from, say, planning a path for a robot arm to avoid obstacles. Luna: And there's research into using knot theory to help robots plan those motions. I remember a paper from a few years ago where they encoded the knot topology and then used search algorithms to find a sequence of moves that would produce it. Lucas: Right, that's the work of the MIT group, among others. But even with a good plan, execution is hard because of the deformability. The thread doesn't stay where you put it; it springs back, it sags, it gets tangled from friction. Luna: So it's not just about knowing the steps; it's about doing them under real-world conditions. And that's where force control and tactile sensing become critical. Lucas: And that's where the latest research is heading. Instead of assuming a rigid world, the robot actively senses the tension and adjusts its actions in real time. It's the difference between a robot that blindly follows a script and one that actually feels what it's doing. Luna: That's a good way to put it. And I think that's the key takeaway for today: the reason robots can't tie knots isn't a lack of strength or speed; it's a lack of sensitivity and adaptability. Once we crack that, it'll unlock a whole new class of dexterous tasks. Lucas: Well said. And there are already interesting attempts to commercialize this. I know a couple of startups are working on surgical robots that can automate suturing, and they're using machine learning to improve their knot-tying reliability. They're not there yet, but they're making progress. Luna: And what about the consumer angle? Could we see a robot that ties our shoes or ties a tie for us? Lucas: Maybe in the long run, but I think the more immediate impact will be in specialized fields like surgery, where the stakes are high enough to justify the cost. For tying a bow tie, you might just have to keep doing it yourself. Luna: Probably for the best. It's a nice ritual anyway. Lucas: For now. So, as we wrap up, let's look forward a bit. What would it take to actually solve knot-tying in a general sense? I think we need better tactile sensors that can be integrated into surgical tools, more robust algorithms for online adaptation, and perhaps a hybrid approach that combines model-based planning with learning-based correction. Luna: And we also need benchmarks so that different research groups can compare their progress. Right now, there's no standard test for knot-tying, which makes it hard to know who's really leading. Lucas: That's a great point. We need a standardized set of ropes, knots, and failure metrics. Without that, it's all anecdotal. Some groups are working on that, but it's early days. Luna: Do you think we'll see a robot tie a knot in a live surgery within the next decade? Lucas: I think it's plausible. The pieces are coming together: better sensors, better algorithms, better hardware. The challenge is integration and safety, which always take longer in medicine. But I'd bet we'll see a fully autonomous knot-tying step in some surgical procedure within ten years. Luna: That's exciting. And it would be a huge win for patients, too, because it could reduce variability and fatigue in surgery. Lucas: Exactly. And that's the bigger picture: it's not about robots replacing surgeons, but about augmenting their skills. Tying the knot is the tedious part; the surgeon can focus on the patient. Luna: Well, I'm glad we finally tackled this one. Next time you see a robot struggling with a shoelace, you'll know why. Lucas: And maybe you'll cut it a little slack. Thanks for listening, and we'll catch you on the next one.