Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Still Can't Screw in a Lightbulb
Transcript
- Lucas: So I screwed in a lightbulb this morning. Took me about three seconds. And I got to thinking — there is no robot on the planet that can reliably do that. Luna: Wait, really? Not even the latest arms from Fanuc or ABB? Lucas: Nope. And it's not for lack of trying. The task sounds trivial — rotate until snug — but it combines three failure modes that robotics labs have been fighting for decades. Thread alignment, torque sensing, and grip fragility. Luna: Alright, let's start with alignment. What's so hard? Lucas: A standard E26 bulb has a screw thread that's about 26 millimeters in diameter. The lead — the distance the thread advances per rotation — is roughly four millimeters. So if the bulb is even a couple of degrees off perpendicular when it first touches the socket, the threads cross. Human instinct is to feel that resistance and back off, wiggle, re-approach. That 'wiggle and push' routine is brutally hard to code. Luna: Because the robot doesn't know whether it's hitting the thread or hitting the rim? Lucas: Exactly. The arm is position-controlled. It knows where it is in space to within a tenth of a millimeter. But when the bulb contacts the socket, the actual forces depend on surface friction, the exact angle of the glass, even the temperature of the metal. A rigid arm just pushes until something bends or breaks. Luna: And this is different from, say, screwing a bolt into a nut — which factories do all day? Lucas: Very different. Bolts are metal-on-metal, they're precisely aligned by jigs, and the robot can apply hundreds of Newtons of force. A lightbulb is glass. You can't apply that kind of force. Also, a bolt's thread is usually started by a human or a feeder — the robot just tightens. Here, the robot has to find the thread itself. Luna: So the second problem is torque. How do you know when it's tight enough? Lucas: Right. Over-tighten and you crack the glass or damage the socket. Under-tighten and it's loose, flickering. A human feels the resistance ramp up and stops. A robot relies on a force-torque sensor at the wrist. Those sensors are good — but they measure forces at the mounting flange, not at the fingertips. The glass itself can flex, the bulb's base can slip inside the gripper. The sensor sees a messy signal. Luna: I remember you talking about USB insertion in episode 26 — similar problem, different scale. Lucas: Exactly. USB plugs also require search and insert with tight tolerances. But a USB plug has a rectangular cross-section, so there's only one correct orientation. A lightbulb thread is symmetric — you can start it at any rotation. That actually makes the search space bigger. And the torque signature for a lightbulb is much lower. We're talking maybe 0.5 to 1.0 Newton-meters. Industrial screwdrivers apply ten times that. Luna: So the robot has to be both precise and gentle. What about the grip itself? You mentioned fragility. Lucas: Most robot grippers are parallel-jaw. They pinch. A lightbulb is a tapered glass dome — pinching it near the base is okay, but the contact area is tiny. Glass is strong in compression, but any shear force from the rotation can cause micro-cracks. Plus, the gripper has to hold the bulb firmly enough to transmit torque, but not so firmly that it shatters. That's a narrow window. Luna: Have any research groups actually solved this? Lucas: There's promising work at MIT's Soft Robotics Lab. They developed a gripper with pneumatic fingers that conform to the shape of the object. It can hold an egg without breaking it, and also screw a lightbulb into a socket. But the cycle time is slow — about fifteen seconds per bulb. And it requires a compressed air supply. Not exactly ready for your home. Luna: Fifteen seconds? I can do it in three. Lucas: And that's the benchmark. For a robot to be useful in a home, it has to match human speed and reliability. Right now, the best research systems succeed about 80 percent of the time. Below 95 percent, it's not a product. It's a demo. Luna: Speaking of demos — is there a deeper lesson here for the robotics industry? Lucas: I think so. The lightbulb problem is a perfect example of why 'general purpose' robot hands are still science fiction. Every time we think we've solved manipulation, a seemingly trivial task humbles us. And it's not about adding more sensors or better vision. Vision can tell you where the socket is, but it can't tell you the thread engagement. That's a tactile problem. Luna: So tactile sensing is the bottleneck. Lucas: It's one of them. High-resolution tactile sensors exist — like the GelSight sensor from MIT — but they're fragile, expensive, and generate terabytes of data per hour. Processing that data in real time for control is a huge compute challenge. And even if you solve that, you still need algorithms that can interpret the data. It's a full-stack problem. Luna: And this is the kind of conversation that makes you appreciate how incredible human hands really are. Lucas: Absolutely. And it's the kind of conversation that we get to have because our listeners support the show. If you find these deep dives useful, consider buying us a coffee — literally. It's buy me a coffee dot com slash fexingo. That support keeps the podcast ad-free and lets us spend time on the weird corners of robotics that most shows skip. Luna: Yeah, and every contribution genuinely helps. We don't run ads, so it's listener-funded all the way. And we're grateful. Lucas: So, back to the lightbulb. One company that's trying to commercialize this is a startup called RightHand Robotics. They focus on piece-picking for warehouses — items like boxes and polybags. But they've also done work on more delicate items. Their gripper uses a combination of suction and pinch. For a lightbulb, though, suction on a curved glass surface is unreliable. Luna: Because of the taper? Lucas: Partly. The suction cup needs a good seal. On a curved dome, the seal is often imperfect, especially if the bulb has a slightly dusty surface. So you'd need to clean the bulb before gripping it. That's one more step, one more failure mode. Luna: So what would a practical home robot need to change a lightbulb? What's the minimal hardware? Lucas: At minimum: a force-torque sensor with sub-0.1 Newton-meter resolution, a compliant wrist that can absorb misalignment, and a gripper with soft, high-friction fingertips. Plus software that can execute a search strategy — small spiral or raster pattern to find the thread. And a torque limit that cuts off before cracking the glass. Luna: That's a lot of hardware for a single task. Lucas: It is. And that's exactly why we don't have robot butlers yet. Each task requires specialized sensing and control. The dream of a single hand that can do everything — flip a pancake, fold a shirt, screw a lightbulb — requires breakthroughs in materials, sensing, and AI all at once. Luna: Any timeline predictions? I know you hate giving those. Lucas: I do hate them. But I'll say this: the research trajectory is encouraging. In the last five years, we've gone from zero reliable demos to about 80 percent success in controlled labs. If that curve continues, we might see a commercial robotic lightbulb changer in maybe ten years. But it'll be a specialized device — not a general-purpose hand. Luna: And it will cost more than a lifetime supply of lightbulbs. Lucas: Exactly. For now, I'll keep using my own hands. And every time I do, I'll be a little more impressed by them.