Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Tie a Knot
Transcript
- Lucas: There is a specific moment in robotics research where normally confident PhD students just stare at the floor and admit defeat. It's the moment someone asks them to program a robot to tie a knot. Luna: A knot. Like tying your shoelaces? Lucas: Exactly. An overhand knot in a piece of rope. Something a four-year-old can learn in a weekend. And yet, as of June 2026, there is no general-purpose robot in the world that can walk up to a random piece of string and tie a secure knot without being told exactly what kind of knot, where the string is, and how stiff it is. Luna: I know we've done episodes about folding towels and peeling potatoes. But knot-tying feels like it should be easier. It's just crossing and looping. Lucas: That's the intuition. But the physics is brutal. A towel is a flat sheet — you can approximate it with a grid of points. A potato is rigid — you can model its shape. A knot involves a flexible, friction-varying, self-occluding string. The robot has to manage contact forces that change the moment the string touches itself. Luna: And it can't see its own fingers? Lucas: That's part of it. Most robotic hands have cameras mounted overhead or on the wrist. When the robot grasps the rope and starts looping, the hand occludes the knot. The visual feedback vanishes at the critical moment. So you're relying on force sensing. And force sensing in most commercial robot arms is terrible — you get maybe one Newton of resolution, which is like trying to feel a feather while wearing a winter glove. Luna: So what separates a knot from, say, a surgical suture? Because there are robots that suture in the OR. Lucas: Surgical suturing is a constrained version. The needle is rigid, the tissue is held in place, the knot is pre-tied or uses a specific instrument. The da Vinci system, for example, doesn't actually tie a free knot — it uses a cinch mechanism. The open problem is a free knot in a deformable rope with no fixtures. That's what the DARPA hackathon in 2023 tried to solve. Luna: The hackathon where nobody succeeded? Lucas: Nobody. Sixteen teams, three days, a single overhand knot. The best attempt got the rope crossed but couldn't pull the loop tight without slipping. It was a humbling reminder that manipulation is the last great barrier in robotics. Luna: And that's with the rope lying flat on a table, right? Not dangling? Lucas: Correct. The rope was pre-positioned, one end clamped. The robot just had to perform the sequence: grasp, loop, pass the end through, and tension. The failure modes were fascinating. Some robots pulled too hard and the knot became a slipknot. Some under-tensioned and the rope just fell apart. One team's gripper crushed the rope fibers, changing the friction profile mid-task. Luna: That reminds me of something — and I realize this is a slight shift, but it ties back. We do a lot of deep dives like this on Fexingo, and we keep the shows free and ad-free. That's possible because a small group of listeners supports us directly. If you've gotten something from this conversation, there's a link — buy me a coffee dot com slash fexingo. It's a simple way to keep these explorations going. Lucas: Yeah, and it genuinely makes a difference. We don't run ads, we don't have sponsors, so it's listener-funded all the way. A coffee a month from a few people adds up. Thanks to everyone who already does. Luna: Alright, back to the knot. So where is the research headed? Because I know there are groups making progress. Lucas: Two main tracks. One is better hardware — specifically, torque-controlled hands with high-resolution tactile sensors. Shadow Robot's Dexterous Hand, for instance, has 24 joints and integrated tactile sensors that can detect slip and texture. They've demonstrated a figure-eight knot with a specific type of paracord. But it's still not general. Luna: The other track is simulation, right? Training in virtual environments. Lucas: Right. Researchers at MIT and NVIDIA have been using reinforcement learning in simulated physics engines — MuJoCo, Isaac Gym — to train policies for knot-tying. They can run thousands of parallel simulations, each with different rope stiffness, friction, and thickness. The challenge is the simulation to real gap. Real rope has nonlinear bending, fiber shedding, and micro-slippage that simulators don't capture. Luna: So you train a policy that works perfectly in simulation, and then it fails on a real rope. Lucas: That's exactly what happened in a 2025 paper from Stanford. Their sim to real transfer rate for a bowline knot was about 12 percent. They had to add domain randomization — randomizing friction, mass, damping in simulation — to get it up to 40 percent. Still not production-ready. Luna: Forty percent success rate on a single knot. That's not even close to reliable. Lucas: And reliability is everything in industry. If a robot fails 60 percent of the time tying a knot, you can't put it on a factory floor. You'd have a pile of tangled cables in minutes. For reference, human workers in cable harness assembly achieve 99.5 percent first-pass yield. So the bar is high. Luna: There's also the question of why we need robots to tie knots at all. Is this just a benchmark, or is there a commercial application? Lucas: Both. The benchmark is valuable because knot-tying tests a whole range of manipulation skills — dexterity, force control, planning under uncertainty. Commercially, there's a huge need in cable harness assembly for aerospace and automotive. A Boeing 787 has over 50 miles of wiring, much of it tied into bundles. That's still done by hand. Also surgical suturing, fishing net repair, textile manufacturing, even tying off trash bags in automated waste handling. Luna: So the market is real. It's just waiting for the technology to catch up. Lucas: Exactly. And there are some promising approaches. One is using magnetic grippers that don't crush the rope. Another is using multiple fingers with independent force control — the Allegro Hand from SimLab has that. A team in Japan used a pair of robotic arms working together, each arm controlling one end of the rope, and they achieved a 70 percent success rate on a square knot. Luna: Two arms? That's cheating a little bit. Lucas: It is, but it's also how humans do it. We use two hands. Most robotic knot-tying research tries to do it with one hand because that's harder. But in practice, a factory might have two arms available. So the dual-arm approach is probably the first commercial viable one. Luna: What about learning from human demonstration? I've seen videos where a human teleoperates a robot hand to tie a knot, and then the robot tries to imitate. Lucas: That's the third track — imitation learning. The problem is that human motion is so fluid that the robot tends to overfit to the specific trajectory. If the rope is two millimeters thicker, it fails. But with enough demonstrations — hundreds of thousands — you can start to generalize. Google's DeepMind did this with a task called 're-grasping,' where the robot had to pick up a pen and rotate it. Knot-tying is that, but on steroids. Luna: It sounds like the fundamental challenge is that robots lack a model of the rope as a continuous, flexible object. They treat it as a series of discrete segments. Lucas: That's exactly right. Most control algorithms assume the object is rigid or has known kinematics. A rope has infinite degrees of freedom. The state space is enormous. So researchers use reduced-order models — like treating the rope as a chain of beads connected by springs. That works for coarse motions but fails for the fine tension adjustment needed in the final pull. Luna: Is there any work on using tactile sensing to feel the knot tighten in real time? Like, feel the torque change. Lucas: Yes, that's exactly what a group at the University of Washington did. They mounted a six-axis force-torque sensor on the wrist of a robot and trained a neural network to predict knot tightness from the force signature. They could classify whether a knot was loose, secure, or over-tightened with 94 percent accuracy. But that was after the knot was already formed — they weren't using the feedback to guide the tying in real time. Luna: So real-time closed-loop control is the next frontier. Lucas: It is. And it's incredibly hard because the sensor data is high-dimensional and the dynamics are fast. A human can adjust tension in milliseconds. A robot control loop running at 1000 hertz might still miss the moment when the rope starts to slip. Plus, you have to distinguish between the force of pulling and the force of the knot settling. Luna: I'm trying to think of a parallel. Is there any industry that has solved a similar problem? Like textile manufacturing? Lucas: The textile industry has automated a lot of things — weaving, cutting, sewing — but knot-tying is avoided. Sewing machines use a lockstitch mechanism that ties a knot between two threads automatically, but it's a specialized mechanism, not a general robot. The robots that handle fabric in garment manufacturing still use grippers that pick up panels, they don't tie anything. Luna: So we're still at the research stage. What's a realistic timeline for a robot that can tie a random knot on a random rope reliably? Lucas: If you asked me in 2020, I would have said five years. Today, I'd say maybe another five to ten. The sim to real gap is closing, but slowly. Hardware is getting better — the new tactile sensors from companies like SynTouch have 10-micron resolution. And foundation models for robotics, like Google's RT-2, are showing some generalization across tasks. But knot-tying requires a level of precision and adaptability that those models haven't demonstrated yet. Luna: It's humbling. We can put a rover on Mars, but we can't tie a knot. Lucas: That's the thing. Space exploration is about surviving extreme environments with known physics. Manipulation is about living in our world with unknown physics. A knot is a microcosm of that challenge. Every time I see a robot fail to tie a simple overhand, I'm reminded how far we are from general-purpose dexterity. Luna: And yet, people keep working on it. It's one of those problems that won't go away until it's solved. Lucas: Exactly. And when it is solved, the impact will be enormous. Not just in factories, but in homes, hospitals, everywhere that needs a knot tied. It's one of those quiet breakthroughs that will unlock a whole new wave of automation. Luna: So maybe by 2030, we'll have a robot that can tie our shoes. But will it also untie them? Lucas: Now you're asking for the sequel episode.