Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Screw in a Lightbulb
Transcript
- Lucas: You know the feeling when you're standing on a chair, one hand holding the lightbulb, trying to feel for the threads and start turning without cross-threading or crushing the glass? That simple motion — screwing in a lightbulb — is a surprisingly difficult benchmark for robotic manipulation. Luna: I can see why. The glass is fragile, the threads require alignment, and you need just the right amount of torque. But I never thought of it as a robotics problem. Lucas: It really is. And it's a task that reveals how much we take for granted in human motor control. The end-effector has to simultaneously apply downward pressure and rotational motion, while adapting to the stiffness of the socket. If the robot pushes too hard, the bulb breaks. If it doesn't push enough, the threads won't catch. Luna: So it's a combination of force sensing and compliance. Sounds like a job for impedance control. Lucas: Exactly. Impedance control allows the robot to adjust its stiffness in response to external forces. But the tricky part is that the required force profile changes throughout the motion. When you first start threading, you need low force and careful alignment. Once the threads engage, you need a firm axial push coupled with smooth rotation. And near the bottom, you need to stop before the bulb is torqued too tightly against the base. Luna: And each socket is different — some are tighter, some have worn threads. It's not a one-size-fits-all solution. Lucas: Right. That's why most industrial robots that handle threaded parts use specialized end effectors and compliant fixtures. They can screw in a bottle cap or a bolt because the geometry is consistent and the material is predictable. But a lightbulb is the wild west: glass that can shatter, a shape that's not perfectly cylindrical, and threads that vary by manufacturer. Luna: Are there any robots that can actually do it reliably? I remember seeing a video from OpenAI's Dactyl hand solving Rubik's cube, but that's in simulation. Lucas: OpenAI's Dactyl did learn to manipulate a cube and even solve a Rubik's cube with a special hand. But that task is about dexterity and finger coordination, not the force-control challenge of threading. For lightbulb screwing, researchers at MIT's CSAIL and a few companies have demonstrated prototypes using tactile sensors. One approach uses a gel-filled fingertip with a camera inside — it captures deformation patterns as the finger presses and twists, giving the robot a sense of contact geometry and slip. Luna: So they're giving the robot a sense of touch. That's starting to sound human-like. Lucas: In a way. Human fingertips have thousands of mechanoreceptors. We can feel the rotation, the friction, the moment the threads catch. A tactile sensor can't replicate that density, but it can provide enough data for a learning-based controller. In one 2024 paper from Stanford, they used a soft gripper with embedded capacitive sensors to screw in a standard household lightbulb with about an 80 percent success rate. Luna: Eighty percent is impressive, but not yet good enough for an automated home or a factory line. Why does it fail the other twenty percent? Lucas: The failures usually come from two places: slippage and misalignment. If the gripper doesn't apply enough normal force, the bulb slips as it rotates. But if it applies too much, the glass can crack, especially if there's a pre-existing stress point. The other failure mode is cross-threading — the bulb starts at an angle, the threads don't engage, and the robot keeps turning, either stripping the thread or damaging the socket. Luna: So it's a delicate balance. And that's just for one task. How do we scale that to the variety of tasks robots need in the real world? Lucas: That's the holy grail of general-purpose manipulation. But for now, researchers are focusing on specific tasks to build up a library of skills. Lightbulb screwing is a great test case because it combines fine force control, visual servoing, and adaptability. Solving it reliably would have implications for other assembly tasks, like inserting a component into a circuit board or attaching a hose coupling. Luna: So it's not really about lightbulbs. It's about developing a control framework that can handle any threaded or insertion task. Lucas: Exactly. And that's why this problem gets a lot of attention in robotics conferences. Every year there are new papers on force-controlled screwing, tactile-based insertion, or learning from demonstration for peg-in-hole tasks. The lightbulb is just a compelling example because everyone has done it, and everyone knows when a robot gets it wrong it's both hilarious and frustrating. Luna: Speaking of which, do you think we'll see a robot that can change a lightbulb in our homes within the next decade? Lucas: Probably not as a standalone feature — it would be one skill among many in a robot assistant. But as a research benchmark, it's already driving progress. And honestly, if we can solve the lightbulb problem, the sensor and control innovations will trickle down to a hundred other applications. Luna: That's the kind of incremental progress that makes industrial automation exciting. And it only happens because people invest in research and development, whether through public funding or private support. Lucas: Speaking of support, if today's look at the mechanics of robot manipulation gave you something useful, you can help us keep having these conversations. It's simple: buy me a coffee dot com slash fexingo. We stay ad‑free and independent because of listeners like you. Luna: No pressure — just if you found value. Every little bit helps us dive deeper into these topics. Lucas: So where does that leave us? The lightbulb task is still unsolved in a general sense, but the approaches are getting more robust. I think the next big leap will come from combining tactile sensing with better material models for the gripper. Soft robotics has a role here, too — a compliant gripper can naturally adapt to some of the variability. Luna: And what about the learning side? Are we seeing robots learn from human demonstration, or is it mostly reinforcement learning from scratch? Lucas: Both. Researchers are using reinforcement learning in simulation to learn a policy for screwing, then transferring it to the real robot. The challenge is the sim to real gap — the physics of glass, thread friction, and compliance are hard to model accurately. Learning from demonstration can provide a good starting point, but it requires a lot of human data. Hybrid approaches seem most promising. Luna: It's a lot like teaching a human apprentice — you show them, let them try, and then correct. Except the robot might need thousands of trials. Lucas: Yeah, and that's why this problem is so fascinating. It sits at the intersection of hardware design, sensing, control theory, and machine learning. Every time we think we've cracked it, a new constraint emerges — like the fact that a cold bulb has different friction than a warm one. Luna: Wait, really? Temperature affects the friction of glass and metal? That's another variable. Lucas: It does. Glass expands slightly with heat, and the metal threads can also change. So a robot that only trained in a 20‑degree lab might fail in a 30‑degree factory. Again, generalizing across conditions is hard. Luna: So the humble lightbulb is really a gateway to understanding the full complexity of robotic manipulation. It's not just a gadget problem — it's a fundamental challenge. Lucas: Exactly. And that's why I think it's worth paying attention to. The day a robot can reliably screw in a lightbulb, we'll know we've taken a big step toward versatile robotic assistants. Luna: And until then, I'll keep doing it myself — carefully. Lucas: Same here. Thanks for listening.