Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Cable
Transcript
- Lucas: So, last week I tried to plug a usb c cable into my laptop while looking away. Just a simple, one-handed, blind plug. Took me about a second and a half. Luna: Yeah, and I'm guessing if a robot tried that, it would either crush the connector or miss entirely. Lucas: Exactly. And that's the topic today: why robot arms still can't handle a cable. Not just plugging things in — picking up a charging cord, routing it through a clip, even just identifying which end is which. Luna: It sounds simple, but I suspect it's anything but. Lucas: It's one of the hardest open problems in manipulation. Think about what a cable is. It's a flexible object with near-infinite degrees of freedom. A rigid block has six degrees — three translations, three rotations. A cable, say a 20-centimeter USB cord, effectively has hundreds of internal bend and twist states. Luna: Right, and every time you move it, those states change. So the robot can't just memorize one shape. Lucas: Exactly. Most industrial robots rely on geometric models. They know the exact dimensions of a part, they grasp it at a pre-planned point, and they move it along a known path. But a cable doesn't hold a fixed geometry. It droops, it coils, it snags on itself. That breaks the entire planning pipeline. Luna: And force-torque sensors don't help much either, right? Because the cable's stiffness varies. Lucas: That's a huge part of the problem. When you push a cable into a port, the resistance isn't constant. It depends on the exact bend radius at the connector, whether the cable is twisted, how much slack there is. A standard force sensor reads 'something is pushing back' but can't distinguish between 'the cable needs a little more force to seat' and 'I'm about to break the port.' Luna: So the robot either under-inserts or over-forces. I've seen videos of robots just smashing connectors into ports. Lucas: Right, and that's with a known cable. Now imagine a robot in a data center, where every cable is a different length, gauge, and stiffness. It's a nightmare. Luna: I read about a startup in Japan that's trying exactly that — automating cable routing in server racks. What's their approach? Lucas: That's a great example. The company is called RoboWired — they're based in Tokyo. They developed a custom gripper with a tactile sleeve that wraps around the cable. The sleeve has 128 pressure points per square centimeter, so it can feel the cable's orientation and stiffness in real time. Luna: So it's not just a rigid clamp. It's more like a hand that can sense the cable's shape. Lucas: Exactly. They claim a 92 percent success rate on routing a cable through a cable management arm — which is a plastic hook thing you see in server racks. That's up from 15 percent with a standard gripper. But the system is slow — about 45 seconds per cable. A human does it in 10 seconds. Luna: Forty-five seconds per cable adds up fast. In a data center with thousands of cables, the labor cost is significant. Lucas: Right, and that's why the economics are tricky. A data center tech costs about $25 an hour. At 10 seconds per cable, that's about 7 cents per cable in labor. But if a robot takes 45 seconds, the per-cable cost — factoring in amortized hardware — is higher. So the robot only makes sense if it can work 24/7 without errors, reducing rework. Luna: Honestly, if today's conversation gave you a new appreciation for just how hard cable manipulation is, and you think that insight was worth a coffee, there's a link — it's buy me a coffee dot com slash fexingo. No pressure, just helps keep the show ad-free. Lucas: Yeah, and we genuinely mean that. Every little bit lets us keep digging into these niche robotics problems that don't get mainstream attention. Luna: So back to the tactile sleeve — 128 pressure points per square centimeter. How does that compare to what's in research labs? Lucas: It's actually on the high end. A recent paper from MIT CSAIL — published in early 2025 — developed a tactile sleeve with 256 points per square centimeter, but it's not as durable. They used it for cable routing tasks and achieved 72 percent success on a complex routing path, compared to 12 percent for a standard gripper. Luna: Seventy-two percent is still not reliable enough for production. You'd need 99.9 percent for a factory. Lucas: Exactly. And that's the big gap. In controlled lab conditions, you can get good results. But add variable lighting, different cable stiffness, dust on the sensor — performance drops. The MIT team noted that their success rate fell to 30 percent when they used a slightly stiffer cable. Luna: So the cable's material properties are a moving target. Is there any work on using computer vision to help? Lucas: Plenty. But vision has its own problems. A cable is a thin, curved object — it's hard to segment from the background. Stereo cameras can estimate the 3D shape, but if the cable is coiled or overlapping, the algorithm gets confused. There's a 2024 paper from Carnegie Mellon that used a neural network trained on 50,000 synthetic images of cables to predict the cable's configuration, and they got about 85 percent accuracy on test images. Luna: Eighty-five percent sounds good, but in real time, a wrong prediction could cause a crash. Lucas: Right. And the sim to real gap is huge. The synthetic images were perfect — no shadows, no occlusion, no dust. In a real data center, the accuracy dropped to 55 percent. So researchers are now combining vision and tactile sensing. The idea is: vision gives a rough guess of the cable's shape, then tactile feedback refines the grip and insertion force. Luna: That makes sense — two modalities compensate for each other's weaknesses. Lucas: Yes. And there's interesting work on using a 'curvature closure' model — basically, the robot uses the tactile sleeve to feel the cable's bending moment and then adjusts its grip to keep the cable straight near the connector. That's critical for insertion, because if the cable is bent near the plug, it won't align with the port. Luna: So the robot has to actively control the cable's shape, not just grab it. Lucas: Exactly. And that requires a control loop running at hundreds of hertz, because the cable's shape changes faster than you'd think. If the robot moves its arm 10 centimeters, the cable might whip and change its bend radius in a few milliseconds. Luna: That's a lot of computation. Are any industrial robot manufacturers tackling this? Lucas: A few. FANUC has a 'cable handling' option on some of their collaborative robots, but it's limited to very stiff cables — think thick power cords, not USB cables. ABB has a research project using their YuMi robot for wire harness assembly, but they haven't commercialized it. The problem is that the market is fragmented: automotive wants wire harnesses, electronics wants USB and HDMI, data centers want patch cables. Each has different requirements. Luna: And the volumes are lower than for, say, picking boxes. So the return on investment for developing a universal cable manipulator is uncertain. Lucas: That's the business reality. But there's a push from the home robotics side. If you want a robot that can clean your desk, it has to handle cables — plugging in a phone charger, untangling headphones. That's a much harder problem because the types of cables are unpredictable. Luna: So the home robot is the ultimate benchmark for cable manipulation. If you can solve it there, you can solve it anywhere. Lucas: Right. And we're not there yet. Boston Dynamics' Spot can navigate stairs, but it can't plug itself into a charging dock reliably. That's partly a cable problem — the charging connector is often on a cable that needs to be aligned and inserted. Luna: I remember a video where a Spot tried to dock and just nudged the cable around. It was almost comical. Lucas: Yeah, it's a reminder that even the most advanced robots have these basic blind spots. And that's what makes this field fascinating — the things we take for granted are the hardest to automate. Luna: So if I were a startup founder, what would be the most promising approach? Tactile sleeves, vision, or something else? Lucas: I'd bet on the combination — vision for global shape estimation, tactile for local force control — but with a focus on low-cost sensors. The MIT sleeve is too fragile and expensive for production. There's a group at UC Berkeley working on a capacitive tactile skin that can be 3D-printed for under $10 per square centimeter. If that scales, it could be a game-changer. Luna: A $10 sensor compared to what, a $2000 force-torque sensor? Lucas: Exactly. And it's flexible, so it can wrap around grippers of any shape. The key is making it durable enough for thousands of cycles. Early tests show it lasts about 500 grasps before the capacitive layers degrade. Five hundred is not enough for a factory, but it's a start. Luna: Five hundred grasps on a cable that's constantly bending and twisting — that's a tough wear test. Lucas: It is. But the research is moving fast. I'd say within three to five years, we'll see the first commercial cable-handling robot that can do a specific task — like plugging in a charging cable — with 99 percent reliability. The general-purpose cable manipulator is probably a decade away. Luna: That feels realistic. And it means for now, humans still have job security when it comes to cable management. Lucas: Definitely. Until a robot can untangle my headphones, I'm not worried.