Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Uncork a Wine Bottle
Transcript
- Lucas: You know, a few months ago we spent an episode on why robot arms still can't open a jar. And we got a surprising amount of listener mail — not just from frustrated home cooks, but from people working in packaging and food manufacturing who said, 'that's exactly our problem at scale.' Luna: I remember that. One person wrote in from a winery, saying if we thought jars were hard, we should try robotically uncorking a wine bottle. Lucas: That's the one that stuck with me. So I spent the last few weeks talking to a team at a research lab in Bordeaux that's actually trying to solve this. And the short version: robot arms still can't reliably uncork a wine bottle. Luna: Wait — is this a real industrial problem? I thought wine bottling was already highly automated. Lucas: It is, but only up to a point. High-speed bottling lines use mechanical corkers that push corks into bottles at like 300 bottles per minute. That part is solved. The hard part is the reverse — pulling the cork out, especially for premium wines that use natural cork, which is... well, a natural material. Luna: So it's variable. Cork density, diameter, friction — all different from bottle to bottle. Lucas: Exactly. The research team I spoke with — they're using a Universal Robots UR10e arm with a custom three-finger gripper and a force-torque sensor at the wrist. Their benchmark is simple: grab a wine bottle, insert a corkscrew, extract the cork without damaging it or pushing fragments into the wine. Do that for 100 different bottles of the same wine, same vintage. Luna: What success rate did they get? Lucas: Initially, 70 percent. Thirty bottles out of a hundred had some kind of failure — the cork crumbled, the corkscrew slipped, or the arm yanked too hard and the cork snapped in half. Luna: That's... not great for a production environment. Even 99 percent failure-free would be a lot of wasted wine at scale. Lucas: Right. And the failures aren't random — they cluster around corks that are either very dense or very brittle. The team measured extraction forces across their test set and found a range from 8 Newtons — that's a pretty easy pull — up to 25 Newtons for a stubborn cork. That's more than a three-to-one variation. Luna: And a robot programmed with a fixed force profile can't handle that spread. It'll either under-pull and stall, or over-pull and break the cork. Lucas: Bingo. So the fix they're working on is adaptive force control. The arm starts pulling with a low target force, but it continuously monitors the actual force feedback. If it detects resistance rising above a threshold, it pauses, increases the target, and continues. It's basically a closed-loop control scheme with real-time adjustment. Luna: That sounds similar to what we discussed with the jar-opening problem — the need for compliance and sensing. Lucas: Very similar. And the results are promising. After implementing that adaptive algorithm, their success rate went from 70 percent to 87 percent on a second batch of 100 bottles. Still not perfect, but a meaningful jump. Luna: Just out of curiosity — what are the failure modes now? The remaining 13 percent? Lucas: Mostly corks that are unusually brittle near the surface. The corkscrew itself — which is a standard steel worm-style screw, not a specialised one — can create micro-cracks when it enters the cork, and if the cork is already weak, that crack propagates during extraction. So now they're experimenting with hollow corkscrews that inject air to create a small cavity before pulling, reducing the surface stress. Luna: So the problem keeps pushing upstream. It's not just the pulling force, it's how you engage the cork in the first place. Lucas: Exactly. And this is where the research stops being about wine and starts being about general manipulation. The same adaptive force control approach could apply to peeling a potato — another thing robot arms can't reliably do, as we've covered — or handling soft medical tissues during surgery. Luna: Quick honest thing — this is exactly the kind of deep-dive episode we can only make because a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo. It's literally listener-funded, no ads, and that's what lets us spend weeks talking to researchers in Bordeaux instead of chasing clicks. Lucas: Yeah, and we're genuinely grateful for that. It keeps the show independent and lets us follow the weird, specific questions. Like why a robot arm still can't uncork a bottle of wine. Luna: Alright, so back to that 87 percent — is that good enough for real-world use? Lucas: For a commercial winery? No chance. A 13 percent failure rate means you'd be throwing away one out of every eight bottles you try to open. But for a research prototype six months in, it's a solid proof of concept. The team expects to hit 95 percent within a year by adding a second sensor — a small camera that visually inspects the cork's surface before insertion. Luna: So pre-screening the cork for cracks or soft spots, and then adjusting the corkscrew insertion depth or angle accordingly. Lucas: Exactly. That multimodal approach — vision plus force sensing — is where a lot of hard manipulation problems are headed. It's not just about having a better gripper; it's about giving the robot enough information to decide how to grip. Luna: And that's fundamentally different from the way most industrial robots work today, which is 'do the same motion 10,000 times and hope the part is consistent.' Lucas: Right. The cork problem is a great example because the 'part' — a natural cork — is inherently inconsistent. You can't tighten your tolerances; you have to adapt. And that's exactly the challenge for robotics in food processing, agriculture, even warehouse picking where items aren't uniform. Luna: So what's the next step for this Bordeaux team? Are they publishing? Lucas: They're aiming for ICRA 2027, the big robotics conference. But they also have a collaboration with a local cooperative winery that wants to use the system for quality testing — they currently open 50 bottles per batch by hand to check for cork taint. If a robot could do that reliably, it would save a lot of wrist strain. Luna: And if it works for wine, it could work for other sealed containers. Maybe even pharmaceuticals, where you need to open sterile vials without contamination. Lucas: That's exactly the kind of cross-industry transfer the researchers are hoping for. The force control algorithms are general; the end-effector is the only part that's wine-specific. Swap the corkscrew for a vial opener, and you've got a new application. Luna: Alright, I'm sold. So when can I buy a robot sommelier? Lucas: Probably not for a few years. But if you're a robotics researcher listening, the Bordeaux team is hiring a PhD student to work on the vision module. And if you're just someone who loves wine — maybe pour yourself a glass and appreciate that your own arm and hand are still way better at this than any robot. Luna: Cheers to that.