Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Drop a Screwdriver Mid-Twist
Transcript
- Lucas: So we've spent over twenty episodes on robot arms failing at everyday tasks. Peeling potatoes, tying shoelaces, folding towels. But here's one we haven't touched, and it's arguably the most industrially important one of all: screwing in a screw. Luna: Really? I mean, you'd think a robot arm would crush that. It's such a repetitive, precise motion. And we're not talking about a delicate task like threading a needle. Lucas: Exactly. And in fact, the auto industry has been trying to automate screwdriving for decades. But the truth is, even today, a huge fraction of final assembly screws are still driven by hand. There's a reason for that, and it comes down to something called torque control. Luna: Torque control. So it's about how much twisting force you apply? That sounds like something a robot could measure pretty precisely. Lucas: You'd think so. But the problem is that torque isn't just a single number you hit at the end. It's a whole curve over the life of the screw turn. You have to manage the ramp-up, the sudden spike when the screw head seats, and then the final tighten without stripping the head or snapping the screw. Luna: So it's not just 'tighten to ten newton-meters', it's 'tighten to ten newton-meters without breaking anything along the way.' Lucas: Right. And that's where it gets hard. Because the physics of a screw joint changes with every material, every thread, every bit of lubricant or dirt on the threads. And those changes are often too fast and too subtle for a standard robot controller to react to. Luna: So what does a human do differently? I mean, I can screw in a lightbulb without thinking about it. Lucas: Actually, that's a perfect example. When you screw in a lightbulb, you're constantly feeling the resistance. You feel when it starts to catch, you feel when it seats, and you adjust your speed and force on the fly. A robot arm with traditional position control doesn't feel anything. It just moves to a programmed position and applies a set current to the motor. Luna: So it's blind. It's like trying to screw in a lightbulb with a blindfold on, but you've memorized the exact number of turns. Lucas: Exactly. And if the lightbulb is even slightly different, or the socket has a different depth, you either strip the threads or you don't screw it in far enough. That's why early automation in auto plants was limited to very consistent, very controlled sub-assemblies. The moment you got to the final line, where every car is a little different, you needed humans. Luna: But that's changing, right? I feel like we're at a point where robots are getting force sensors and better algorithms. Lucas: It is changing, but slowly. And the key breakthrough has been adding a force-torque sensor to the robot's wrist. That lets the arm feel the resistance as it turns. But sensing is only half the battle. The harder part is the control software that can interpret those forces in real time and decide what to do next. Luna: And that's the part that's been really hard to get right, right? Because the forces can change so quickly. Lucas: Yeah. Think about a typical screw joint. As you turn, the torque rises gradually, then there's a sudden jump when the screw head meets the surface. That jump might happen in a few milliseconds. If the controller doesn't react fast enough, you overshoot and either strip the head or snap the screw. Luna: So it's like a split-second decision. And a human does that naturally, but a robot has to be programmed to anticipate it. Lucas: Right. And the early solutions were basically to slow the robot down to a crawl near the end, so it didn't overshoot. But that killed cycle time. And in a factory, that's money. So the challenge has been to keep speed while adding enough intelligence to react safely. Luna: And that's where the modern stuff comes in. There are robots now that can do this at speed, right? I've seen videos of these cells that just drive screws all day long. Lucas: Right. And the trick is something called impedance control, or admittance control. Instead of just following a position path, the robot treats its own arm like a spring. It can give a little when it meets resistance, and it can push harder when it needs to. That's how a human arm works, actually. We're not stiff; we're compliant. Luna: So the robot learns to be compliant. But isn't that a trade-off? If the arm is too springy, it might not have the rigidity to drive a screw straight. Lucas: Exactly the tension. You want rigidity for the approach and the initial thread catching, but compliance for the final seat. And the good systems switch between the two on the fly, based on what the force sensor is telling them. Luna: That sounds like a really fine line to walk. And I imagine the programming is not trivial. Lucas: It's not. Which is why, for a long time, the best practice in many factories was to use a specialized screwdriving tool mounted on a robot, rather than the robot itself. The tool would have its own motor and its own torque sensor, and the robot would just position it. But that adds cost and complexity. Luna: And you still have the issue of alignment. The screwdriver bit has to be perfectly coaxial with the screw, or you'll cam out and strip the head. Lucas: That's a whole other problem. And it's one that's been getting a lot of attention lately, because even with a force sensor, if the bit is off-axis by even a couple of degrees, you'll round out the screw head or damage the workpiece. Luna: So how are people solving that? I've seen some research on using vision to locate the screw head and then aligning the tool to it. Lucas: Vision helps, but it's not enough by itself because the screw head might be recessed or the lighting might be inconsistent. So a lot of systems now use a combination of vision for gross positioning and then a mechanical compliance in the tool, like a floating head, that lets the bit self-center as it touches the screw. Luna: So it's like the robot gets close, and then the tool does the fine alignment. That makes sense. Lucas: And that's why you're still seeing a lot of hybrid systems. But the exciting recent work is in using machine learning to improve the torque control itself. Instead of hand-coding the torque curve, you train a model on thousands of screwdriving cycles, and it learns to recognize the subtle force signatures that precede a successful seat or a failure. Luna: So the robot learns from experience. And I guess that means it can adapt to different screw types and materials without being reprogrammed. Lucas: Exactly. And that's the promise of the next generation of screwdriving robots. There are startups now that are building plug and play screwdriving cells that can be retrained for a new product in hours, not weeks. That's a game-changer for small manufacturers who have mixed production runs. Luna: And speaking of small manufacturers, I've been thinking about the economics. A robot arm with a force sensor and the software to do this is not cheap. When does it actually pay off? Lucas: That's the million-dollar question. Literally. A decent six-axis robot with a wrist force-torque sensor and the control software can run you anywhere from fifty thousand to a hundred and fifty thousand dollars, depending on the payload and reach. And then you have to integrate it, program it, and maintain it. Luna: So for a small shop that's doing a few hundred assemblies a day, that might be hard to justify. Lucas: Right. But the equation changes if you're a large manufacturer doing millions of screws a year. And it's not just labor savings. It's consistency. A robot doesn't get tired at the end of a shift, and it doesn't have bad days. Plus, with the data from the force sensor, you can do real-time quality control, flagging any screw that doesn't meet spec. Luna: That's a good point. Consistency is often undervalued in these cost-benefit analyses. But it can be huge in terms of warranty claims and recalls. Lucas: And that's actually where I think the future is going. We're not just talking about screwdriving, but about robots that can feel their way through any assembly task. The force-sensing wrist is like giving a robot a sense of touch, and that opens up a whole new category of tasks that were previously impossible to automate. Luna: It's funny, because we started this series about all the things robots can't do, and now we're seeing that with the right sensors and software, they're starting to do them. Lucas: Yeah. And honestly, a lot of that progress is thanks to listeners who support the show. If today's episode gave you something useful, and you want to keep these deep dives coming ad-free, the best way is to buy us a coffee. It's a small thing, but it genuinely keeps the lights on. Luna: And it's a nice feeling to know you're part of making that happen. Lucas: The link is buy me a coffee dot com slash fexingo. And now, back to the machines. Because the next frontier is not just screwing, but actually handling flexible parts, like cables and hoses, which is a whole other nightmare. Luna: Oh, I bet. But let's save that for another episode. For now, I'm just impressed that we're finally at a point where robots can screw in a screw without stripping it. Lucas: It took fifty years, but we got there. And the next fifty are going to be even more interesting.