Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Twist a Screw
Transcript
- Lucas: There’s a moment on every assembly line that you never think about until it goes wrong. A robot arm picks up a screw, moves it to a threaded hole, and begins to turn. And then — nothing. The screw doesn’t seat. The robot keeps applying torque, the threads strip, and suddenly you’ve got a multi million dollar line stalled because a robot can’t finish the last quarter-turn of a screw. Luna: Wait — this is a robot we’re talking about, right? The same six-axis arm that can weld a chassis to within a tenth of a millimeter? A screw is that hard? Lucas: Harder than almost anything else in assembly. And it’s not just one robot — I’ve been looking at data from a German automotive supplier, let’s call them PrecisionTech GmbH, who run about 400 robots across three plants. Their engineers told me that screw-insertion failures cause roughly 12 percent of their unscheduled downtime. That’s more than vision failures, more than gripper failures. The humble screw. Lucas: And look — We deliberately don't run ads on these shows. If you want to support that choice, the link is buy me a coffee dot com slash fexingo. That simple gesture helps us keep the conversation ad-free and focused on the engineering. Luna: Yeah, it's a small way to keep the show exactly as it is. No sponsors, no scripts — just us digging into what actually breaks in a factory. Lucas: Right. So back to the screw. The problem starts with something called chamfer detection. When a human tightens a screw, their wrist naturally finds the taper at the mouth of the threaded hole — the chamfer — and guides the screw in. The robot’s controller doesn't have that tactile intuition. It knows its joint positions to within a few arc-seconds, but it can't feel the chamfer the way we do. Luna: So the robot is essentially blind to the geometry before the threads engage. It just pushes the screw straight down until it hits something. Lucas: Exactly. And if the screw’s approach angle is off by even half a degree, the tip skids across the chamfer, the controller registers a force spike, and it either aborts the operation or keeps turning and cross-threads the hole. That’s how you get a stripped bolt and a rework ticket. Luna: I’ve seen those rework tickets. They’re not cheap. A single cross-threaded hole in an engine block can add twenty minutes of manual repair, and the line is down the whole time. Lucas: Right. So the industry has tried to solve this with force-torque sensors at the wrist. Those sensors measure the six-degree-of-freedom forces and torques the robot feels at the tool flange. The idea is that the controller can detect the lateral forces when the screw tip hits the chamfer and then servo the arm sideways to center it. Luna: Does that actually work in practice? I’ve seen demos where it looks flawless — the screw just glides right in. Lucas: Demos are one thing. Production is another. The problem is friction. A force-torque sensor can tell you the magnitude and direction of the force, but it can't distinguish between the force from contacting the chamfer and the force from, say, a burr on the workpiece or a slightly oily surface. The noise floor is high. And the control loop has to run at kilohertz rates to be reactive enough. Luna: So the sensor is good, but the signal processing isn't there yet to tease apart the actual contact event from environmental noise. Lucas: That's the engineering challenge. PrecisionTech tried a machine-learning approach where they trained a neural network on thousands of successful and failed insertions. They gave the network the six-axis force-torque stream plus the robot's joint velocities and currents. And it improved success rates from about 93 percent to 97.5 percent. That sounds great, but three failures in a hundred means you're still getting a stoppage every few minutes on a high-speed line. Luna: And that 2.5 percent failure rate is probably concentrated on the hardest cases — like when the screw is slightly misaligned in the gripper or the hole has a tiny chip. Lucas: Exactly. The easy 93 percent are the ones where the tolerances are perfect. The remaining failures are the edge cases. And in a factory, edge cases are the ones that cost you the most downtime because they're unpredictable. Lucas: There's another approach that a few research labs are exploring: using reinforcement learning to train the robot to 'feel' its way in. Basically, you let the robot try thousands of insertions in simulation, with a reward function that penalizes high forces and rewards successful seating. Then you transfer that policy to the real arm. Luna: But sim to real transfer for contact-rich tasks is notoriously hard. The friction coefficients in the real world are never exactly what you modeled. Lucas: Right, and the robot's own compliance — the slight flex in its joints and links — changes how the forces propagate. A simulation that assumes a rigid arm will overestimate the contact forces. The real robot deflects a bit, which changes the geometry. So the policy learned in simulation can fail when it hits the real hardware. Luna: So where does that leave us? Are we going to have robots that can reliably screw things together anytime soon? Lucas: Probably not in the next two or three years for the highest-tolerance applications. But there's a promising hybrid approach: use a vision system to roughly align the screw, then a force-controlled search strategy that sweeps the screw in a small spiral until it drops into the hole. That's what some of the newer collaborative robots are doing. It's slower than a human, but it's more reliable. Luna: Slower but reliable — that's the trade-off. And for low-volume, high-mix assembly, that might be good enough. But for high-volume automotive, you need speed and reliability. Lucas: And that's why humans are still on those lines. The robot can do 90 percent of the work, but the last 10 percent — the tricky fastening, the cable routing, the snap-fit — that's still a person. The screw is a great example of how far we've come, and how far we still have to go. Luna: It's also a reminder that the hardest problems in robotics aren't the flashy ones. It's the mundane, high-precision, high-reliability tasks that we take for granted. Lucas: Exactly. And that's why I find this episode so interesting. Next time you tighten a screw by hand, just think: that simple motion is the frontier of industrial robotics.