Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Still Can't Handle a Screwdriver
Transcript
- Lucas: You know how IKEA furniture comes with that little Allen key? Imagine a robot trying to use it. That's actually a surprisingly deep problem in robotics — and it's why most assembly lines still need humans for anything involving a screwdriver. Luna: I've watched warehouse robots pick boxes and palletize cans, but I've never seen one drive a screw. Why is that so hard? Lucas: Let's start with the physics. When you torque a screw, the tool applies a rotational force, and the robot's wrist has to counter that torque. But unlike a human wrist, which has passive compliance — our muscles can yield slightly — most robot arms are rigid. They fight the torque, which can lead to stripped threads or the screw slipping out. Luna: So it's not just about knowing where the screw goes. It's about managing the forces during the turn. Lucas: Exactly. There's a great paper from MIT's CSAIL from 2024 — they tested a Franka Emika Panda arm on a standard screwdriver insertion task. The robot failed about 30 percent of the time when the screw had to go into a tight hole with a slight angle mismatch. Humans? They do it without thinking, because we can feel the resistance and adjust instantly. Luna: Thirty percent failure is not production-ready. What's the bottleneck — sensing or control? Lucas: Both, but the sensing part is wild. Most robot arms have joint torque sensors, but those measure torque at the motor, not at the tool tip. So there's a whole chain of gears and linkages that introduce friction and backlash. The robot doesn't actually feel what the screwdriver tip is doing. It's guessing based on a noisy signal several joints away. Luna: So you'd need a sensor right at the wrist, or even at the tool tip. Lucas: Right. Some research groups are experimenting with strain gauges embedded in the tool itself. But that adds cost and complexity. And even if you have perfect sensing, the control problem remains: how do you coordinate the linear push into the screw with the rotational motion? Too much push and you jam the screw; too little and it doesn't catch the threads. Luna: It sounds like a coordination problem — like patting your head and rubbing your stomach, but with force feedback. Lucas: That's a great analogy. And unlike the human brain, which has a dedicated cerebellum for fine motor coordination, robots have to solve this with algorithms. Most use impedance control — they model the arm as a spring-damper system and try to maintain a desired stiffness. But those models break down when the tool contacts the surface at an angle. Luna: I'm curious — is this just an academic problem, or does it matter in industry? Lucas: It matters a lot. Think about final assembly in automotive: doors, dashboards, interior trim — all screwed together. Or electronics: smartphones and laptops have dozens of tiny screws. Right now, those tasks are almost entirely manual. If you could automate them, you'd save billions in labor costs and improve consistency. Luna: So the incentive is huge. Why hasn't anyone solved it yet? Lucas: Partly because the easy problems — pick and place, palletizing, welding — were solved first, and they captured low-hanging fruit. Screwdriving is a high-precision, high-variability task. Every screw is slightly different: different torque specs, different alignment, different materials. Automating it requires a system that can adapt on the fly. Luna: And that's where machine learning comes in, I assume. Lucas: Yes, but it's not just throwing data at it. Researchers are using reinforcement learning to train policies that can handle the torque feedback. A team at UC Berkeley trained a robot to drive screws into a board with varying densities — soft wood, hard wood, particle board — and it learned to modulate its speed and force based on the resistance. But it took thousands of episodes, and the robot still struggled with screws that had a bit of debris in the thread. Luna: So it's fragile. One speck of sawdust and the policy breaks. Lucas: Exactly. And that's the challenge with learned policies: they're amazing in the training distribution, but they fail in edge cases. The real world is full of edge cases. A slightly worn screwdriver bit, a screw with a burr, a hole that's been painted over — all of those break the model. Luna: Which brings us back to the sensing problem. You need to detect those anomalies in real time. Lucas: Right. Some startups are working on what they call 'force-guided assembly.' They use a six-axis force-torque sensor at the wrist and run a control loop that looks at the instantaneous torque profile. When the robot drives a screw, the torque should follow a known curve: low at first, then a spike as the screw seats, then a final peak when it's tight. If the curve deviates, the robot backs out and tries again. Luna: That's basically what a human does when they feel the screw start to cross-thread. Lucas: Exactly. And it works — for simple cases. But it's slow. Each retry adds a few seconds, and in a high-volume assembly line, that adds up. Plus, the robot has to know what the 'correct' torque profile is for every screw in every material. That's a lot of tuning. Luna: What about tool design? Could you make a screwdriver bit that helps the robot? Lucas: There's interesting work on that. Some labs are using compliant bits — like a spring-loaded sleeve that allows a bit of angular misalignment. That passive compliance can absorb up to 5 degrees of error. It's like having a universal joint in the screwdriver. But it also reduces the torque you can apply before the spring bottoms out. Luna: So there's always a trade-off. Lucas: Always. And that's why this problem is still open. It's a classic robotics trade-off: precision versus speed versus robustness. You can optimize two, but not all three. Luna: You know, this kind of reminds me of something: a couple of dollars a month from listeners is genuinely what keeps shows like this going — buy me a coffee dot com slash fexingo, if you've gotten something out of these deep dives. Lucas: Yeah, it makes a real difference. And it helps us keep the show independent and ad-free. So if today's tech conversation gave you something usable, that's the way to support it. Luna: Anyway — back to screws. One thing I wonder: could we just redesign screws to be more robot-friendly? Lucas: That's a great point. Some manufacturers are moving toward screws with a tapered lead-in, which makes alignment easier. And there's a push for standardized torque specs across industries. But changing decades of screw standards is like turning a battleship. Still, it's happening gradually. Luna: So in the meantime, robots will keep struggling with our IKEA furniture. Lucas: For now, yes. But I think we'll see a breakthrough in the next five years. The combination of better force sensors, more sophisticated control, and perhaps even soft robotic wrists could make screwdriving routine. And when that happens, it will unlock a huge swath of automation in assembly and construction. Luna: I'd love to see a robot build a whole IKEA dresser. Lucas: There's actually a lab at MIT that did that — but it took them 20 minutes per piece, and they had to use custom-designed parts. So we're not there yet. But it's a fascinating benchmark for dexterity. Luna: What's the next step you're watching for in this space? Lucas: I'm keeping an eye on tactile sensing skins — thin sensors that can wrap around robot fingers and measure shear forces. If you can feel the screwdriver handle slipping in your grip, you can adjust. That's something humans do unconsciously. A tactile skin could give robots that same feedback. Luna: So it's not just about the tool — it's about the hand holding the tool. Lucas: Exactly. And that's why this problem is so rich. It touches on sensing, control, materials, and even design. It's not just a screwdriver; it's a window into the whole challenge of robotic manipulation. Luna: Well, next time I'm wrestling with a stubborn screw, I'll remember that robots have it worse. Lucas: That's the spirit. And maybe by the time we need to assemble our next bookshelf, a robot will be able to help.