Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Untie a Knot
Transcript
- Lucas: So here's a question that I think gets at the heart of why robotics is still so hard: when was the last time you watched a robot untie a knot? Luna: I don't think I've ever seen that. I've seen robots tie knots — there's that old video of a robotic arm tying a surgical suture. But untying? Lucas: Exactly. And there's a reason for that. Tying a knot is actually a fairly structured task. The robot knows the sequence — loop, pull, tighten. Untying, though, requires the robot to understand the current state of the knot, which could be any configuration of a twisted, tangled string under unknown tension. Luna: So it's not just about dexterity. It's about perception and reasoning. Lucas: Right. And this is where we get into what roboticists call 'the knot problem.' A knot is a topological object — its properties don't change if you stretch or bend the string. But to untie it, you need to know exactly which strand goes over which, and apply forces in the right directions without making it tighter. Luna: And a robot arm with a standard gripper has almost no ability to feel that. It can't sense the tension in the string, or whether it's about to slip. Lucas: That's the core issue. Most industrial robot arms operate in what's called 'position control' — they move to a precise point in space, with a certain force, and they assume the object is where they expect it to be. But a string doesn't stay put. It's flexible, it deforms, it has nonlinear behavior. Luna: Nonlinear meaning the relationship between force and movement isn't straightforward. A little pull might do nothing, then suddenly the whole knot tightens. Lucas: Exactly. So untying a knot requires what roboticists call 'reactive manipulation.' The robot has to sense the state of the knot in real time and adjust its grip and motion accordingly. And that's extremely difficult with current tactile sensors. Luna: Let's talk about those sensors. What's state of the art? Lucas: So the best tactile sensors today are things like the BioTac, which is a fingertip-shaped sensor that can measure pressure, temperature, and even micro-vibrations. Or more recently, MIT's GelSight, which uses a camera and a gel pad to reconstruct the shape of what's being touched with high resolution. Luna: But even those struggle with something as thin as a string. The contact area is tiny, and the sensor might not capture the full geometry. Lucas: Right. And there's a latency issue. By the time the robot processes the tactile data, decides what to do, and moves, the knot may have changed. So researchers are working on faster control loops — some are getting down to the millisecond range. Luna: Is there any robot that can actually untie a knot reliably today? Lucas: Yes, but only in very controlled settings. Last year, a team at MIT's CSAIL published a paper on a robot that can untie knots in ropes. They used a dual-arm setup with GelSight sensors and a planning algorithm that models the knot's topology. Luna: How did it work? Lucas: So they start with the robot observing the knot with a camera. Then it uses a technique called 'topological mapping' to identify the crossing points — which strand is on top. Then it plans a sequence of manipulations: it might use one arm to hold the knot steady while the other arm pulls one end to loosen it. Luna: And it worked? Lucas: It worked on several types of knots — overhand, figure-eight, even a bowline — with success rates around 80 percent. But each knot took about 10 minutes. And if the rope was too thin or too slippery, it failed. Luna: So still a long way from a robot that can untie your shoelaces in the morning. Lucas: A very long way. And that's partly because shoelaces are a different material — they're flat, they have high friction, and they're often tied in a bow, which is a more complex topological structure. Luna: I want to ask about the broader implications. If robots can't untie knots, what does that mean for tasks like surgery or manufacturing? Lucas: Well, in surgery, there's a lot of knot-tying for sutures. But untying is less common. However, in manufacturing, there are tasks like cable harnessing — where robots have to manipulate bundles of wires — that are still almost entirely done by humans because wires get tangled. Luna: So the knot problem is actually a bottleneck for automation in certain industries. Lucas: Absolutely. And it's not just knots. It's any task that involves flexible, deformable objects. Think about folding clothes, handling food, or even assembling a smartphone — there are cables and connectors that need to be routed. All of those involve the same fundamental challenge. Luna: So what's the path forward? Better hardware or better software? Lucas: Both. On the hardware side, we need more dexterous hands with higher-resolution tactile sensing. Some labs are working on soft grippers that can conform to objects — like the ones from Soft Robotics Inc. — but they lack precision. On the software side, we need better models of how flexible objects behave, and faster algorithms to plan manipulations. Luna: There's also the learning approach. Some groups are using reinforcement learning to teach robots to untie knots by trial and error. Lucas: That's promising, but it requires a huge amount of data — and in simulation, the physics of strings is hard to model accurately. So you train in sim, then transfer to the real world, and it often fails because reality is messier. Luna: Right. It's the sim to real gap. And for something like a knot, the gap is especially wide because friction and tension are so sensitive. Lucas: And that sensitivity is why a human can untie a knot without thinking — we have proprioception, we can feel the tension, we subtly adjust. A robot has none of that instinct. Luna: It makes you appreciate how remarkable even a toddler's dexterity is. A two-year-old can untie a knot if they try. Lucas: And that's a humbling benchmark. A toddler's brain and hand are far more advanced than any robot system we've built. So when we ask why robot arms can't untie a knot, the honest answer is: because we haven't figured out how to give them the senses and the intelligence that a human child has. Luna: I think for listeners who are following this from a business or technology perspective, the key takeaway is that this isn't just a niche problem. It's emblematic of a whole class of manipulation tasks that are still unsolved. Lucas: Exactly. And if you found that useful — if today's episode gave you a clearer picture of what's actually hard in robotics — I just want to mention that the way we keep this show ad-free and focused on the real engineering challenges is through listener support. You can help at buy me a coffee dot com slash fexingo. Luna: Yeah, it's a small thing that makes a big difference for us. And we appreciate everyone who does. Lucas: So back to the knot problem. One recent development that I think is really interesting: a team at Stanford used a dual-arm robot with a novel 'tension-aware' controller. Instead of trying to model the knot explicitly, they had the robot gently pull on the rope while monitoring the force, and then release when tension spiked, like a human would. Luna: So it's a more reactive approach — less planning, more feedback. Lucas: Exactly. And they got good results on simple knots. But it still doesn't generalize to new knot types or different materials. Luna: It sounds like the field is moving from trying to solve knot untying as a planning problem to treating it as a control problem. Lucas: I think that's right. And that shift — from model-based to model-free, from open-loop to closed-loop — is happening across many areas of robotics. But for knots, the challenges are particularly acute. Luna: Let's look ahead. If someone wanted to work on this problem, where should they start? Lucas: I'd say start with understanding topology and knot theory — there's a beautiful mathematical foundation. Then get into tactile sensing and control. And if you're a software person, look at reinforcement learning with simulators like MuJoCo or Isaac Sim. Luna: And for companies? What's the commercial opportunity? Lucas: Huge. Any factory that deals with cables, wires, or textiles could benefit. The global wire harness market alone is over $50 billion, and a lot of that assembly is manual. If you can automate even 10 percent of it, that's a big win. Luna: So the robot that can untie a knot might not just be a lab curiosity — it could be a transformative piece of industrial hardware. Lucas: That's the hope. But we're probably still years away from a commercial product. In the meantime, every time you untie a shoelace, give your fingers a little credit. Luna: I will. And maybe I'll stop complaining when my robot vacuum gets tangled in a charging cable. Lucas: Now you know why it happens. Thanks for listening to The Robotics Podcast.