Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Open a Jar
Transcript
- Lucas: So, you're home, you've just made a sandwich, you reach for the pickle jar, and — twist. Nothing. Twist again, harder. Still stuck. You grab a dish towel for grip, give it a whack on the counter, and finally — pop — the lid breaks loose. Congratulations. You just did something no industrial robot arm can reliably do. Luna: Open a jar? Seriously? Lucas: Seriously. And I'm not talking about some fancy multi million dollar humanoid robot. I'm talking about the standard six-axis robot arm you'd see on a car assembly line or in a warehouse. Give it a pickle jar with a twist-off lid, and it will fail more often than it succeeds. This is episode 106 of the show, and we've spent over a hundred episodes on what robot arms can't handle — paper, trash, eggs, feathers, deflated balloons. But jars are a special kind of nightmare. Luna: Okay, I need to understand this. Why is a jar harder than, say, a bowling pin? We already covered that one. Lucas: Great question. A bowling pin is rigid. You can grip it with a parallel-jaw gripper, apply a known force, and pick it up. A jar lid requires two things at once: high torque to break the seal, and precise control to not crush the glass or strip the lid. And the robot has to do both without any of the tactile feedback you used — the feel of the lid slipping, the sound of the seal breaking, the angle of your wrist. Luna: So it's a sensing problem, not just a strength problem. Lucas: Exactly. Let's put some numbers on this. A typical pickle jar lid requires about 5 newton-meters of torque to open. That's roughly the same as tightening a bolt with a short wrench. But here's the catch: the robot's wrist torque sensor has to measure that in real time, while also maintaining a grip force that's strong enough to prevent slipping but weak enough not to crack the glass. And the lid itself — those little ridges? They're designed for human fingers, not robot grippers. Luna: So what happens when a robot tries? Does it just spin forever? Lucas: Typically one of three things. Option one: the gripper doesn't get enough purchase, so it just slips — the arm rotates, the lid doesn't. Option two: the gripper crushes the glass because it's applying too much normal force while trying to generate torque. Option three: it successfully opens the lid, but then can't pull it off cleanly — the lid tilts and jams. Luna: And this is with a standard industrial arm, right? Not something designed for jars? Lucas: Right. A standard arm like a Fanuc or KUKA. They're built for repetitive, precise motions with known geometries. A jar lid is a variable geometry — the diameter changes slightly between jars, the lid material flexes, the seal strength varies. It's not a known, fixed part. And the robot has no equivalent of your palm or fingertips to sense slip. Luna: I remember from the egg episode — the problem was similar. Too much force and you break it. Too little and you drop it. But a jar adds rotation. Lucas: That's the key addition. With a linear pick and place, you're controlling force in one axis. With a jar, you need simultaneous control of radial grip force and tangential torque. And the ratio between them changes as the lid loosens. Early in the twist, you need high torque and high grip. Once the seal breaks, the torque drops sharply — but you still need grip to unscrew the lid all the way. Most robot control algorithms don't handle that dynamic transition well. Luna: So what's the state of the art? Are there robots that can do it, even if not perfectly? Lucas: Yes, but they're research prototypes, not commercial products. There's a well-known paper from MIT's CSAIL lab — 2023, I think — where they used a soft robotic gripper with embedded tactile sensors. The gripper is basically a set of flexible fingers filled with a granular material that can stiffen on command. It conforms to the lid shape, senses slip via changes in electrical resistance, and adjusts grip in real time. They got a success rate of about 85 percent on a set of 20 different jars. Luna: Eighty-five percent is impressive, but not production-ready. Especially if you're trying to do it on a fast-moving line. Lucas: Exactly. And the other approach is vision-based. Some labs use cameras to track the lid's rotation and estimate slip optically. But that adds cost and complexity, and it's still slower than a human. A skilled factory worker can open a jar in under a second — 800 milliseconds, maybe — and have a 99.9 percent success rate. The robot is at maybe 2 seconds and 85 percent on a good day. Luna: So why does this matter beyond the pickle jar at home? Where would a robot that can open jars actually be useful? Lucas: Surprisingly many places. Food processing plants where jars of sauce, jam, or pickles need to be opened for quality control. Pharmaceutical labs where vials with twist-off caps are used. Even recycling centers — imagine a robot that can separate a glass jar from its metal lid. But the deeper point is that jar-opening is a benchmark for a broader capability: dexterous manipulation of non-rigid, variable-geometry objects with threaded fasteners. Luna: Threaded fasteners — so it's not just jars. It's any screw-top container. Bottles, fuel caps, chemical drums. Lucas: Right. And that's a huge category in manufacturing and logistics. If you could teach a robot to reliably open any threaded container, you'd unlock automation in dozens of industries. But right now, we're stuck at the pickle jar. Luna: It's funny — we've had robots on Mars, robots performing surgery, but opening a jar is still a frontier. Lucas: Well, the Mars rover doesn't have to open a jar of Martian pickles. But honestly, the problem is a perfect illustration of Moravec's paradox: things that are easy for humans — like sensory-motor coordination — are hard for robots, and things that are hard for humans — like complex math — are easy for robots. Jar-opening is squarely in the easy for humans, hard for robots category. Luna: So what's the next step? Are we close to a commercial solution? Lucas: A few startups are working on it. One that caught my eye is called SoftGrip — they're developing a universal gripper with integrated torque sensing and adaptive compliance. They claim a 95 percent success rate on a standard jar, but they haven't published third-party validation yet. And there's also work at UC Berkeley on using reinforcement learning to teach a robot to open jars by trial and error — essentially letting the robot bang the lid on the counter like we do. That's actually a promising approach because it doesn't require perfect sensing — just a policy that learns to exploit friction and impact. Luna: Wait, the robot literally smacks the jar on the table? That's kind of brilliant. Lucas: It is. The idea is that by tapping the lid on a hard surface, you can break the seal without needing high torque. That's exactly what we do, right? We hit the lid on the counter to deform it slightly and break the vacuum. So the robot learns to do the same — it taps, then twists. The Berkeley group got to about 90 percent success with that method. But it's still slower than a human, and you need a hard surface nearby. Luna: So the solution might not be better sensing, but better strategy. Like, instead of trying to mimic human strength, mimic human tricks. Lucas: Exactly. And that's where I think the field is heading — away from brute-force control and toward behavioral strategies. We've had a hundred episodes on what robots can't do, and the pattern is clear: the hardest problems are the ones that require adaptive, tactile, multi-modal intelligence. Jars are just the latest example. Luna: You know, this conversation has been genuinely useful. It makes me think about all the little things we take for granted. Lucas: If today was worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. It's the smallest ask, and it keeps this show ad-free and focused on exactly these kinds of deep dives. Luna: Yeah, a small way to support a big curiosity. Totally optional, but appreciated. Lucas: So back to the jar. The thing that strikes me is how this problem connects to a bigger trend in robotics: the shift from rigid, position-controlled arms to compliant, force-controlled systems. For decades, industrial robots were designed to be stiff — you want a welding arm to hit the same spot every time. But that stiffness is exactly the problem when you need to conform to a jar lid. Luna: So the next generation of arms needs to be softer, more adaptive, more like — I don't know — a human arm? Lucas: Not exactly like a human arm, but with variable stiffness. There's a concept called 'series elastic actuators', where a spring is placed between the motor and the joint. That gives you inherent compliance — the arm can 'give' a little when it encounters resistance. That's what you need for jar opening. And those actuators are already being used in collaborative robots — cobots — but they're not yet widespread in heavy industrial settings. Luna: And the sensors? What's the state of tactile sensing for robot hands? Lucas: It's improving fast. The most promising technology is called 'tactile skin' — arrays of pressure sensors that can be wrapped around a gripper finger. They measure force distribution in real time, which lets the robot detect slip and adjust grip. Companies like GelSight and SynTouch make these. But they're expensive — a single fingertip sensor can cost thousands of dollars — and they're still not as robust as human skin. Luna: So cost is a barrier. But if you're opening jars in a factory, you might be willing to pay for reliability. Lucas: Exactly. And the volume matters. A typical food processing plant might open tens of thousands of jars per shift. If the robot can match human speed and reliability, the ROI is huge — you save on labor costs and ergonomic injuries. Carpal tunnel from opening jars is a real thing. Luna: I never thought of that. So there's a genuine business case here, not just a robotics demo. Lucas: Absolutely. In fact, I spoke with a production engineer at a major pickle company — off the record — and he told me they've been trying to automate jar opening for years. They've tried vacuum grippers, custom end-effectors, even a pneumatic system that tries to blow the lid off. Nothing works reliably. So they still have humans doing it. That's the reality. Luna: So we're still at the point where the best robot for opening a jar is a person. Lucas: For now. But the research is accelerating. I'd bet within five years we'll see a commercial robot arm that can open a jar with 99 percent reliability. The pieces are all there — soft grippers, tactile sensing, reinforcement learning. It's just a matter of integration and cost reduction. Luna: And then maybe we can retire the 'robot arms can't open jars' episode and move on to something even harder — like peeling an orange. Lucas: Don't give the producers ideas. But seriously, that's exactly the trajectory. Each solved problem reveals a harder one underneath. That's why this show will never run out of material.