Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Lightbulb
Transcript
- Lucas: So here's a thing that's been bugging me. You know how every few months some robotics company puts out a video of a robot arm doing something impressive — flipping a pancake, stacking boxes, maybe even folding a towel. Luna: Yeah, and it usually takes a hundred attempts to get that one perfect take. Lucas: Right. But there's a task that I've never seen a robot do reliably, and I think it's because it combines almost every hard problem in manipulation at once: screwing a lightbulb into a socket. Luna: Huh. That sounds simple. I mean, humans do it without thinking. Lucas: We do. But think about what's actually happening. Your hand has to align the threads — which are a helix, not a straight line — with the corresponding helix inside the socket. You have to apply just enough axial force to engage the thread, then switch to a rotational torque that's high enough to drive it home but low enough not to crack the glass. And the whole time, your wrist is passively compliant, automatically adjusting for any misalignment. A robot arm typically has zero compliance unless you explicitly program it. Luna: So it's the combination of force and position control in a tight loop. That's hard for industrial robots because they're usually built for stiffness. Lucas: Exactly. A typical six-axis arm on a factory floor has a repeatability of maybe plus or minus 0.02 millimeters. That sounds precise, but it's precision in free space. When you make contact with the lightbulb — especially if the socket isn't perfectly aligned — that same precision becomes a liability. The arm will push straight ahead instead of following the thread, and either strip the threads or shatter the bulb. Luna: I've seen videos of robots trying to screw in bulbs and they just... crunch. Glass everywhere. Lucas: Yeah, it's brutal. So the research community has been working on this. One of the more interesting approaches comes from a group at UC Berkeley. They're using reinforcement learning to teach a robot to adapt its grip and torque in real time based on tactile feedback from the fingers. They call the system FRIDA — it's a soft robotic hand with force sensors embedded in the fingertips. Luna: Soft robotics seems like the obvious direction here. You need that compliance. Lucas: It does. In one of their experiments, FRIDA successfully screwed in sixty-watt bulbs about seventy percent of the time after training. But when they switched to a different bulb — same thread size, different glass shape — the success rate dropped to forty percent. The model had overfitted to the specific geometry. Luna: So even with learning, generalisation is still the bottleneck. Lucas: It's the same problem that shows up in every manipulation task: the real world has too much variation for current learning algorithms to handle robustly. And that's why most industrial robots still avoid fragile objects altogether. I talked to an engineer at a warehouse automation company last month — they handle boxes and totes, but they explicitly forbid their arms from touching anything that could break. It's just not worth the downtime. Luna: That makes sense from a business perspective. But it also means the robots are confined to very structured environments. Lucas: Exactly. And that's fine for factories and warehouses. But if we ever want robots to help in homes, they're going to need to handle a lightbulb. Or a wine glass. Or a raw egg. Luna: Which we've covered in past episodes. Lucas: We have. But the lightbulb problem is a good lens because it combines threading, fragility, and alignment. It's like a benchmark for dexterous manipulation. Luna: Let's talk about the hardware side. What sensors would you need to make this work reliably? Lucas: You'd need a six-axis force-torque sensor at the wrist, which adds cost and weight. Then you'd need tactile sensors on the gripper fingers — capacitive or piezoresistive — to detect slip and contact forces. And you'd need a control loop that runs at least a kilohertz to react fast enough. Right now, that sensor suite alone costs around ten to fifteen thousand dollars per arm, which is more than many small businesses are willing to spend on a robot that changes lightbulbs. Luna: So the economics don't work yet, unless you're in a niche like aerospace or semiconductor manufacturing where precision is worth the premium. Lucas: Exactly. And even then, you'd still have the software problem. You can't just buy a sensor and plug it in. You need algorithms that can interpret the data and generate commands in real time. That's where most of the research is focused. Luna: There's also the question of the object itself. Lightbulbs come in different shapes, sizes, thread pitches. Some are coated with a lubricant, some aren't. Some are frosted glass, some are clear. Lucas: All of which changes the friction, the grip point, the torque required. It's a nightmare for a vision-based system because the visual appearance doesn't tell you the physical properties. Luna: I remember a paper from MIT where they used a 'grasp exploration' strategy — the robot would try a few gentle squeezes to estimate the object's hardness and friction before attempting the full task. Lucas: That's promising. But it adds time. In a factory, every second counts. So there's a trade-off between reliability and speed that the industry hasn't resolved. Luna: And for a home setting, speed might not matter as much. You'd rather have a robot that takes thirty seconds to change a bulb than one that breaks it in two. Lucas: Right. So the home environment might actually be the place where this gets solved first, because the constraints are different. You don't need a cycle time of three seconds. You need robustness. Luna: It's interesting that the same problem looks very different depending on whether you're solving for a factory or a living room. Lucas: Yeah, and that's a theme we come back to a lot on this show: the context determines the solution. For the lightbulb problem specifically, I think we'll see a hybrid approach in the next few years — a stiff industrial arm with a soft, sensorised end-effector that can switch between high-speed pick and place and delicate manipulation. Luna: So basically a robot with two modes: brute force and gentle touch. Lucas: Exactly. Some companies are already working on that. There's a startup called Soft Robotics Inc. that makes grippers with inflatable fingers, and they've demonstrated grasping eggs and lightbulbs. But they haven't yet solved the threading part — they still rely on a human to align the bulb first. Luna: That's a big limitation. If the robot can't do the alignment, then it's not fully autonomous. Lucas: It's not. And that's where the research community is pushing. One approach is to use visual servoing with a camera mounted on the wrist, so the robot can see the socket and align the bulb visually before making contact. But then you need good lighting and a clean view, which isn't always available. Luna: And if the socket is in a ceiling fixture, the camera angle is awkward. Lucas: Exactly. So some groups are exploring acoustic sensing — listening to the sound of the threads engaging to detect whether the alignment is correct. That's still very early stage. Luna: There's also the possibility of using magnetic fields. If you embed a small magnet in the bulb, the robot could sense the orientation. Lucas: That's clever, but it adds cost and complexity to the bulb, which is supposed to be a cheap consumable. I think the real breakthrough will come when we combine cheap sensors with better algorithms, not when we redesign the object. Luna: So where does that leave us in mid-2026? Are there any commercial products that can change a lightbulb reliably? Lucas: Not that I know of. There are some research demos that work in lab conditions, but nothing you could buy and trust to do it every time in your home. The closest thing might be the robots used in semiconductor cleanrooms, which handle fragile wafers — but those are extremely expensive and operate in highly controlled environments. Luna: It's a good reminder that some of the most basic human tasks are the hardest to automate. Lucas: It really is. And I think that's a healthy perspective to have when we see flashy robotics demos online. The mundane stuff — changing a bulb, folding a fitted sheet, peeling a banana — that's where the real frontier is. Luna: And if today's tech conversation gave you something useful to think about, we appreciate that a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo. It's what helps us keep making deep dives like this without ads or sponsors. Lucas: Yeah, it's a small community that makes this possible, and we're grateful. Anyway, back to the robot that can't change a bulb — I think the next big step will be when someone combines a force-torque sensor with a learning-based control policy that can adapt to different bulbs on the fly. That's probably a PhD thesis or two away. Luna: So maybe in episode seventy we'll have an update. Lucas: I hope so. Until then, if you see a robot trying to screw in a lightbulb, maybe stand back. Luna: And maybe have a broom handy for the broken glass. Lucas: Right. Thanks for listening, everyone — we'll be back next week with another deep dive into the hardest problems in robotics.