Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Grippers Can't Handle a Leather Belt
Transcript
- Lucas: So here's a test for you: take a leather belt — like a dress belt — and pick it up off a table with a robot arm. Easy, right? Well, not really. It turns out robot grippers are surprisingly bad at handling limp, high-friction objects like a leather belt. Luna: Wait, I've seen robots pick up car doors and engine blocks. A belt should be a snap. Lucas: You'd think. But car doors are rigid and predictable. A belt flops, it folds, it slides — and the surface friction is actually a huge problem. Leather has a high coefficient of friction — around 0.3 to 0.5 against dry metal — but it's also flexible. So when a gripper tries to grab it, the belt deforms, the contact patch shifts, and the grip fails. Luna: So the gripper's sense of touch — force feedback — can't compensate for the material squirming around? Lucas: Exactly. Most industrial grippers use parallel jaws with force sensors. They close until they hit a target force. But with leather, the belt might buckle inward, or the jaw might just push it away. The sensor reads a force spike, thinks it's good, but the belt is barely pinched. Then when the arm moves, the belt slips out. Luna: That feels like a problem we've solved with soft robotics — compliant grippers that conform to the object. Lucas: You'd think so, but soft grippers have their own issues. They're great for delicate things like fruit, but leather is both delicate and tough. A soft gripper might not apply enough force to hold it securely, especially if the belt is polished or has a slick finish. And if you increase pressure, you risk damaging the leather. Lucas: I actually talked to an engineer at an automotive plant last month. They have a station where a robot has to pick a leather belt — a timing belt, not a dress belt — from a conveyor and place it onto a pulley. The robot failed about 15 percent of the time. The fix? They added a strip of sandpaper to the gripper pads. That boosted friction just enough. Luna: Sandpaper as a hack — that's surprisingly low-tech for a multi million dollar assembly line. Lucas: And it works. But it shows how far we are from truly general-purpose grasping. Now, the thing is — this isn't just about belts. It's about any object that combines high friction with low stiffness. Belts, straps, cables, fabric, even certain plastics. And here's where it gets interesting: the real bottleneck isn't hardware, it's physics and control. Luna: Hard to write an algorithm for something you can't model easily. Lucas: Right. The deformation of a flexible object is nonlinear. Finite element models are too slow for real-time control. So robots rely on trial and error, or human-designed jigs. And that's why you still see humans handling belts in most factories. Luna: If today gave you something to think about — like why your car's timing belt is probably still put on by a person — the show stays ad-free because of listener support. If that's useful, you can buy us a coffee at buy me a coffee dot com slash fexingo. Just a simple way to keep these conversations going. Lucas: Yeah, it really helps. And we don't do ads, so it's just you and the topic. Anyway — back to belts. So one promising direction is what's called 'active friction modulation'. Some labs are working on grippers that can change their surface friction in real-time using electrostatic adhesion or even tiny vibrations. Luna: Electrostatic adhesion — like sticking a balloon to a wall? Lucas: Pretty much. Apply a voltage, and the gripper pad generates an attractive force. It works on leather, fabric, even rougher surfaces. But it's still early — the grip strength is modest, and it doesn't work well if the surface is dusty or oily. Luna: So for now, sandpaper and human hands are winning. Lucas: For now. But there's another approach: using vision to predict how the belt will deform before the gripper even touches it. Some researchers at MIT trained a neural network on thousands of belt-grasping attempts. It predicts the optimal grasp point and angle. They got success rates up from 60 percent to 85 percent. Luna: That's not bad, but it's not 99.9 percent either. And in a factory, even 5 percent failure is huge. Lucas: Exactly. And that reliability gap is why this problem persists. Let me give you a concrete number: in a typical automotive plant, a robot that handles rigid parts might fail once in ten thousand picks. For flexible parts like belts, the failure rate can be one in a hundred. That's two orders of magnitude worse. Luna: So the economics are brutal. You can't replace a human if the robot needs constant monitoring. Lucas: Right. And the cost of a failure isn't just the dropped part — it's the line stoppage, the damage to the part, the safety risk. So many plants simply keep a human at that station. Luna: What about using multiple grippers? Like one to hold the belt flat, another to pick it up? Lucas: That's done sometimes — dual-arm robots or cooperative grippers. But it increases complexity and cost. And it still doesn't solve the friction issue on the gripping surface itself. Lucas: There's also a clever trick from a startup I visited: they use a vacuum gripper but with a porous foam pad that conforms to the belt. The vacuum pulls the belt against the foam, and the foam's compliance helps distribute the force. That worked for flat belts, but not for twisted or folded ones. Luna: So the variety of ways a belt can lie on a surface — flat, twisted, bunched — each requires a different strategy. Lucas: Exactly. And that's the core challenge: general-purpose manipulation. We have robots that excel in structured environments — pick a rigid part from a fixed location, place it precisely. But the real world is messy. And leather belts are a perfect example of that mess. Luna: It's almost philosophical. We think of robots as perfect, but they're incredibly fragile in the face of physical nuance. Lucas: And that's why I find this so interesting. The last couple of decades gave us incredible progress in sensing and control, but we still can't match a human's ability to just feel the object and adjust on the fly. A human picks up a belt, their fingers automatically adjust pressure, angle, and grip type — all subconsciously. Luna: That's the thing — we don't even think about it. But for a robot, it's a sequence of complex decisions with incomplete data. Lucas: Right. And the data problem is huge. A typical force-torque sensor gives you six axes of data. But a human hand has thousands of tactile receptors. Replicating that density is still impractical for industrial use. Luna: So where does that leave us? Five years from now, will robots be handling belts on the line? Lucas: I think we'll see incremental improvements. Better simulation models, more robust vision-guided grasping, and maybe specialized grippers for each material class. But a single robot that can pick up any belt, any orientation, any condition? That's still a research problem. Maybe a decade out. Luna: It's humbling. For all our AI hype, a leather belt is still a frontier. Lucas: It really is. And that's the story of robotics: the harder problems aren't the ones that look hard. They're the ones that look easy. Next time, we'll look at another deceptively simple task — and why robots still can't handle a chain.