Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Thread a Cable Through a Wall
Transcript
- Lucas: So we've spent a hundred and forty-five episodes asking why robot arms can't do the things humans find trivial. Today I want to talk about one that genuinely surprised me when I read the paper: why robot arms still can't thread a cable through a wall. Luna: Wait, like a USB cable through a hole in a desk? That's the thing where you have to blindly wiggle it around until it catches? Lucas: Exactly that. Or an ethernet cable through a wall plate, or a wiring harness through a firewall in a car. It's a task that every human who's ever set up a home theater has done, usually with some choice words. And robots are still terrible at it. Luna: I was just about to say, I've done that with a coat hanger and a flashlight. How hard can it really be? Lucas: That's the perfect setup, because the answer is fundamentally hard. A team at Carnegie Mellon's Robotic Manipulation Lab published a study last year where they had a dual-arm robot try to thread a flexible cable through a hole in a panel. Forty trials, and the robot succeeded eleven percent of the time. Luna: Eleven percent. So worse than random, basically. Lucas: Random isn't great either, but yeah. And it's not because the robot is dumb or the sensors are cheap. It's because a cable is a flexible object, which means it has essentially infinite degrees of freedom. Luna: So unlike a rigid block or a metal cylinder, it can bend, twist, buckle, and sag. There's no single shape it's in. Lucas: Right. And the robot has to control the tip of the cable, but it can only grasp the base. The tip is maybe twenty centimeters away, and its exact position is never fully known, because the cable is flexing and bouncing as it moves. Luna: That's why we've had episodes about soldering and cutting, but this feels even more chaotic. With a rigid object, you know exactly where the tip is relative to the gripper. Lucas: Exactly. With a rigid object, the kinematics are trivially solved. But a cable is a continuous elastic body, and the math for that is far more complex. The standard models are called Cosserat rod models, and they try to capture bending and torsion, but they're computationally heavy and still don't perfectly predict real-world behavior. Luna: Plus, the wall hides the tip. So the robot can't see where it is once it starts going in. It has to rely on force feedback and maybe a little bit of vision at the entry point. Lucas: And in the Carnegie Mellon study, they gave the robot both. It had a wrist camera and a force-torque sensor at the gripper. But the problem is that when the cable hits the back of the wall, the force signal is tiny, and it's easy to mistake it for a snag or a bend. Luna: So the robot pushes harder, and the cable just buckles somewhere in the middle. You've got a loop forming behind the wall, and the robot has no idea. Lucas: That's the bucking problem, and it's one of the core reasons the success rate is so low. The robot's model of the cable doesn't account for buckling well, because the cable is so flexible that even small forces cause it to deform in ways that are hard to predict. Luna: And this isn't just an academic exercise. This is the kind of task that comes up in construction, automotive assembly, and even in surgery. Threading a suture through tissue or a catheter through a vessel is the same underlying challenge. Lucas: Sure. And in surgery, the stakes are a bit higher than a home theater. But the physical principles are the same. That's why so many labs are working on it. Luna: So what's the state of the art? How are people trying to crack this? Lucas: There are a few promising directions. One is tactile sensing. Instead of just a force-torque sensor at the wrist, you put high-resolution tactile sensors on the fingers, so the robot can feel the cable sliding and twisting in real time. That gives you a much richer signal. Luna: So it's not just 'there's resistance,' but 'the cable is rotating slightly and slipping to the left.' Lucas: Exactly. And researchers at MIT have shown that with that kind of feedback, a robot can follow a cable along its length, which is a prerequisite for threading. But even with tactile sensing, the success rate for threading is still not great. Luna: What about reinforcement learning? I feel like every robotics problem these days is being thrown at a neural network. Lucas: And with some reason. The Carnegie Mellon team actually trained a reinforcement learning policy in simulation, then transferred it to the real robot. That's how they got the eleven percent, which is actually better than the non-learning baseline of zero percent. Luna: So the learning is helping, but it's not there yet. What's the bottleneck? Lucas: The sim to real gap. In simulation, you can have perfect state estimation, perfect sensing, and no noise. The real world has cable sag, static friction, and slight variations in cable stiffness. So the policy trained in a perfect world doesn't transfer perfectly. Luna: And then there's the human factor. When I thread a cable, I use my other hand to guide the tip, or I bend the cable to make it stiffer, or I twist it to help it align. Lucas: That's a huge point. The Carnegie Mellon robot was a dual-arm system, so it could use one arm to hold the cable and the other to guide it. But the coordination between the two arms is itself a hard problem. The two arms need to figure out how to share the flexible object without fighting each other. Luna: It's almost like a multi-agent problem, where each arm has to predict what the other is doing. Lucas: Right. And some researchers are exploring a more human-like approach: instead of trying to feed the cable straight in, the robot deliberately bends the cable into a stiffer shape, like a slight curve, so it's easier to control. That's a clever trick, but it requires even more dexterity. Luna: It's funny, you know. We've talked about folding towels, peeling potatoes, snapping green beans. And now threading a cable. These are all things that humans do without thinking, and robots just can't. Lucas: And it's not just about the arm doing the pushing. The perception problem is just as hard. The robot has to estimate the cable's shape in real time, which means tracking a continuous curve from camera images. That's a lot harder than detecting a rigid object's pose. Luna: So you're saying that even if the arm were perfect, the robot might still fail because it can't see what it's doing. Lucas: Exactly. And that's why some teams are adding extra sensors, like a second camera from a different angle, or even a laser line scanner to reconstruct the cable's 3D shape. But that adds cost and complexity. Luna: So where does this leave us? Are we going to have robots doing this in the next five years? Lucas: For controlled environments, maybe. But for messy, unstructured real-world scenarios, I think we're still a decade away. And I think that's okay, because it's a genuinely hard problem. Luna: It's interesting, because we've been doing these episodes for a while, and the listener support is part of what keeps us going. If you've gotten something out of these deep dives, a couple of dollars a month at buy me a coffee dot com slash fexingo really makes a difference. Lucas: Yeah, it genuinely does. That's what keeps these episodes ad-free and lets us spend the time digging into papers like this one. Anyway, back to the cable problem. Lucas: One thing that gives me hope is the idea of learning from human demonstration. Some labs are having humans teleoperate the robot through the task, collecting data, and then training the policy on that. That way, the robot learns the human's tricks, like bending the cable a certain way or using the other arm to guide. Luna: So you're saying the robot is essentially learning to cheat like we do. Lucas: Exactly. And that's a reasonable path forward. But even with that, the robot needs a lot of demonstrations to generalize to different cable stiffnesses, different hole sizes, different wall thicknesses. And that's a lot of data to collect. Luna: And I guess that's the key takeaway: this problem is a combination of modeling, sensing, and control, and there's no single magic solution. Lucas: Right. The one thing I'd want listeners to take away is that when you see a robot fail at something like this, it's not because the robot is dumb or the engineers are bad. It's because flexible object manipulation is one of the hardest open problems in robotics. And it's at the intersection of physics, perception, and learning. Luna: And maybe the next time you're struggling to thread a cable behind your TV, you can feel a little better knowing that even the most advanced robots are right there with you. Lucas: That's the silver lining. Thanks for listening, everyone. We'll be back next week with another puzzle.