Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Handle a Zipper
Transcript
- Lucas: So I tried to get my winter coat zipped this morning, one-handed, while holding coffee in the other. And I thought — a robot arm would absolutely fail at this. Luna: Did you actually try it with a robot, or is this just a thought experiment? Lucas: Thought experiment. But I am not exaggerating. Zipping a zipper is genuinely one of those tasks that seems trivial to us but is a nightmare for a robot. And there's actually a new paper out of MIT CSAIL from earlier this year — February 2026 — that tried to teach a robot arm to zip a duffel bag. Luna: Okay, I need to hear the success rate. Lucas: After 500 training episodes in simulation, the robot succeeded about 40 percent of the time. And that was on a duffel bag laid flat on a table — not a coat on a moving person. Luna: Forty percent? That's… not great. But I guess it's better than zero. What makes it so hard? Lucas: Let's break down the zipper mechanics. A zipper has two rows of interlocking teeth, a slider that rides up to mesh them, and a stopper at the top. The robot has to first align the two halves — and they have to be parallel within maybe a millimeter or two. Then it has to insert the pin into the slider at the right angle. Then it has to pull the slider up with just enough force to engage the teeth, but not so hard that it jams or pulls the fabric. Luna: And that force changes partway through, right? Because the resistance increases as more teeth mesh. Lucas: Exactly. The force profile is nonlinear. And the robot doesn't have tactile sensing anywhere near human-level. Most industrial grippers just have binary contact sensing — touched or not touched. They can't feel the subtle tension buildup the way we can. Luna: So the CSAIL team — did they use a special gripper, or just a standard robot arm? Lucas: They used a Franka Emika Panda arm — that's a common research robot with seven degrees of freedom and force-torque sensing at the wrist. So it can measure forces in three axes and torques in three axes. But even with that, the robot couldn't reliably feel the moment when the slider starts to engage the teeth. The force signature is too subtle. Luna: I've seen videos of robots doing much more complex assembly tasks — like inserting a peg into a hole, which requires sub-millimeter precision. Why is a zipper harder than a peg-in-hole? Lucas: Great question. Peg-in-hole is essentially a single insertion along one axis. The geometry is symmetric, and you can use search strategies to find the hole. A zipper involves multiple sequential insertions along a curved path, with varying friction and alignment requirements at each tooth. And the fabric is flexible, so the parts don't stay fixed — they shift as you pull. Luna: So it's not just a manipulation problem; it's also a perception problem. The robot has to track the position of the zipper in real time. Lucas: Exactly. In the CSAIL experiment, they used an overhead camera and AprilTags on the zipper to track its pose. But even with visual feedback, the robot struggled with the insertion step. They found that the biggest failure mode was the pin missing the slider — it would bump against the side and then get stuck. Luna: So what's the state of the art now? Has anyone solved it completely? Lucas: Not really. There's some work from the University of Tokyo using a soft gripper with tactile sensors that can detect the zipper teeth pattern. They got up to about 70 percent success on a stationary zipper — like on a pillowcase. But on a moving target, like a jacket on a person, it drops back down. And nobody has demonstrated reliable zipping on a garment worn by a human. Luna: And this matters for more than just robot fashion. There are real applications — assistive robotics for people with limited mobility, automated garment manufacturing, even space suits where astronauts have to zip up in bulky gloves. Lucas: Absolutely. The garment industry is still hugely labor-intensive. Zipper insertion is one of the last steps in a lot of apparel assembly that hasn't been automated. And for assistive robots, being able to help someone dress is a big deal. The CSAIL paper explicitly mentions that as a motivation. Luna: I wonder if the solution is better hardware or better software. It seems like both are needed. Lucas: I think it's both. On the hardware side, you need higher-resolution tactile sensing — maybe a biomimetic fingertip like the ones from SynTouch or the GelSight sensor from MIT. On the software side, you need better models of fabric dynamics. Current physics simulators don't handle cloth-tooth interactions well. The friction coefficients are hard to calibrate. Luna: So we're still years away from a robot that can help a paralyzed person put on a jacket. Lucas: Probably. But the progress is real. Five years ago, the success rate was near zero. Now it's 40 percent on a flat duffel bag. That's a meaningful step. And the field is getting more attention because of the broader push for general-purpose dexterous manipulation. Luna: You know, speaking of getting more attention — one way our listeners can help us keep exploring these specific, quirky robotics challenges is by supporting the show. If you've gotten something out of today's deep dive into zipper mechanics, consider buying us a coffee at buy me a coffee dot com slash fexingo. It helps keep the podcast ad-free and focused on the details. Lucas: Yeah, that's right. We don't have a big production team — it's just the two of us and a lot of research. And listener support really makes a difference. Now, back to zippers — I wanted to mention one more thing from the CSAIL paper. Luna: What's that? Lucas: They tried a learning from demonstration approach. They had a human teleoperate the robot to zip the bag a few times, and then the robot tried to imitate. That actually improved performance a lot — from 40 percent to about 60 percent. So there's promise in combining imitation learning with reinforcement learning. Luna: So if we can just show a robot how to do it a few times, it can learn? That sounds a lot like how humans learn — we watch someone, then try, then fail, then adjust. Lucas: Exactly. The difference is that humans have millions of years of evolved sensorimotor control. Robots are starting from scratch. But they learn fast. The question is whether they can generalize — can a robot that learned to zip a black duffel bag also zip a red nylon jacket? Probably not yet. That's the next frontier. Luna: So we're at the stage where robots can kind of zip, but only in very controlled conditions. It's like a toddler learning to do up their own coat. Lucas: Exactly. And just like a toddler, they'll get better with practice. But it might take a few more years before a robot can zip your jacket without you having to hold still. Luna: I'll keep practicing my one-handed zipping in the meantime. Lucas: Good plan. And we'll keep tracking the progress. Next time, we might look at buttons — which is another nightmare entirely.