Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Uncork a Wine Bottle
Transcript
- Lucas: So I tried to open a bottle of wine last night — and failed. The cork crumbled, got stuck, and I ended up pushing it into the bottle. Luna: That is a classic frustration. But I have a feeling you didn't bring that up just to share a kitchen mishap. Lucas: You know me too well. Because I started thinking: if I, a human with opposable thumbs and decades of experience manipulating objects, can mess that up — what would a robot do? Luna: Probably turn the bottle into shards. But is there actual research on this? Lucas: There is. And it turns out uncorking a wine bottle is a surprisingly deep robotics problem. We've done episodes on zippers, spaghetti, wet towels — all hard in their own way. But cork extraction hits a unique set of challenges. Luna: Alright, walk me through it. What makes a cork so fiendish? Lucas: First, the cork is a deformable, compressible, brittle object with high friction. When you push a corkscrew into it, the cork wants to crumble, not stay intact. Then you have to pull it out — which requires a carefully controlled force profile. Too little force, the cork doesn't budge. Too much, and it tears or the bottle slips. Luna: And the robot has to manage that without crushing the bottle or spilling wine. That's a lot of sensory feedback. Lucas: Exactly. One of the key papers on this comes from the Berkeley Robot Learning Lab. They set up a Franka Emika Panda arm — a seven-degree-of-freedom collaborative robot — with a two-finger gripper and a special corkscrew tool. The goal was to have the robot autonomously uncork a bottle. Luna: So a seven-axis arm with force-torque sensing — that's a high-end setup. Did it work? Lucas: Partially. The robot could successfully insert the corkscrew about seventy percent of the time. But the extraction phase was where it fell apart. The arm would apply a constant upward force, but the cork would start to deform asymmetrically, and the gripper would lose alignment. In about half the trials, the cork broke or the bottle rotated. Luna: So even with state of the art force control, the robot can't handle the subtle adjustments a human makes instinctively. Lucas: Right. Humans use tactile feedback from the entire hand — not just fingertip force. We sense the cork starting to twist and we adjust our grip angle. We can also vary the pulling speed based on resistance. Most robots today rely on joint torque sensors, which give a delayed, filtered version of that feedback. Luna: So it's a sensing problem as much as a control problem. But couldn't you just add more sensors? Lucas: You could, and researchers have tried. Some groups use tactile skin arrays, like the ones from SynTouch, that measure pressure distribution. But then you have to integrate that data into the control loop in real time — and that's computationally intensive. Plus, the cork surface is irregular, so the sensor data is noisy. Luna: And a human doesn't think about any of this. They just pull. Lucas: Exactly. Which is why this is a benchmark task for compliant manipulation. If a robot can uncork a bottle reliably, it probably has the dexterity and sensing to handle a lot of other delicate tasks — like assembling electronics or handling lab equipment. Luna: Speaking of lab equipment — I've seen some startup demos that claim to automate wine service. Are they using robots or just clever mechanisms? Lucas: Most of the commercial 'robotic wine openers' are really just motorized corkscrews in a fixed frame. They work because the bottle is clamped, the cork is centered, and the path is predetermined. That's not general manipulation. The robot we're talking about has to handle a bottle placed arbitrarily on a table, with no fixture. Luna: So the challenge is in the unstructured environment. And that's the frontier of robotics right now. Lucas: Quick honest thing — a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that's literally what funds making this many episodes like this. Luna: Yeah, it's a small group, but it keeps the show ad-free and lets us dive into these weird, specific problems. Lucas: So if today's tech conversation gave you something usable, that's the reason. Luna: Back to the cork. So what's the next step? Can we train a robot to get better at this? Lucas: Yes, and that's where reinforcement learning comes in. The Berkeley group ran thousands of simulated cork-pulling trials in MuJoCo — a physics simulator. They gave the robot a reward for successfully extracting the cork without breaking it. After about ten thousand simulated attempts, the policy learned a smoother, more adaptive pull strategy. Luna: Ten thousand attempts — that's a lot of virtual wine. Lucas: Right. And when they transferred that policy to the real robot, the success rate jumped from seventy percent to about ninety percent for insertion, and extraction success went from fifty percent to seventy-five percent. Still not perfect, but a big improvement. Luna: What about the remaining failures? Are they all due to cork variability? Lucas: Mostly, yes. Natural corks have varying density, grain, and moisture content. Synthetic corks are more uniform, but they're also more slippery. The robot's gripper had trouble maintaining a secure hold on synthetic corks — it would just slip off. Luna: So the material science of the cork itself is a variable. That's a tough problem to generalize across. Lucas: Exactly. One approach is to use a different end-effector — like a three-fingered hand instead of a two-finger gripper. The third finger could provide lateral stability. But then the control complexity goes up. Luna: It seems like every solution adds new challenges. Is there any commercial application for this, or is it purely academic? Lucas: There are commercial implications. If a robot can handle a cork, it can handle many other deformable or fragile objects — like medical vials, glassware, or food items. Companies like DexNet have shown that learning-based grasping works for a wide range of objects, but the real bottleneck is still the manipulation after the grasp. Luna: So the cork problem is a microcosm of the broader manipulation challenge. Lucas: Exactly. And the lesson is that we're still far from human-level dexterity. But each small failure — like a crumbled cork — teaches us something about compliance, sensing, and control that applies to the next generation of robotic hands. Luna: I'm never going to complain about opening a bottle myself again. Lucas: I feel the same way. Even my failed attempt last night — it's a reminder that our own hands are extraordinary machines. And building an artificial equivalent is one of the hardest engineering problems of our time.