Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Screw
Transcript
- Lucas: So, we've done a lot of episodes on why robot hands can't handle soft, slippery, or delicate objects. But there's a basic assembly task that still trips up most industrial robot arms, and it's something a human can do in under a second without thinking. Luna: Let me guess — driving a screw. Lucas: Exactly. Driving a screw. It sounds trivial, but it's actually a nightmare for a robot. And I don't mean the task of picking up a screw and placing it — that's been solved for decades. I mean the actual act of rotating the screw into a threaded hole, applying the right torque, and stopping at exactly the right tightness without stripping the threads or snapping the head off. Luna: And without the robot arm just shaking itself apart from the reaction forces, right? Lucas: Right. Because when you drive a screw, the torque you apply doesn't just go into the screw — it twists the entire robot arm. Most industrial arms are designed for position control: they move to a specific coordinate in space. But screw driving requires force control — you have to maintain a precise axial force while simultaneously applying rotational torque. And the dynamics change as the screw seats. Luna: So the robot has to 'feel' when the screw bottoms out, or when it's cross-threaded. That's a sensing problem. Lucas: It is. And the standard approach — using a compliant wrist or a passive torque limiter — works for simple applications, but not for high-precision work. Think about a smartphone assembly: there are dozens of tiny screws, each with a specific torque spec, often in awkward angles. Most factories still use human workers for that step. Luna: I've read that in automotive assembly, screw-driving stations are some of the hardest to automate fully. Even with collaborative robots, they often end up with a human overseeing the final torque check. Lucas: Yeah, and that's where a recent paper from MIT's CSAIL lab caught my attention. Published late last year, a team demonstrated a robotic system that could drive screws into a standardized test board with 92 percent success rate across different screw sizes and materials. The key innovation was in the control loop: they used a force-torque sensor at the wrist and an algorithm that switched between three modes — search, drive, and seat — based on real-time feedback. Luna: Three modes — so the robot isn't just doing one motion, it's adapting. Lucas: Right. In search mode, it's finding the hole — because even with computer vision, there's always a tiny offset. Then drive mode applies a constant axial force while rotating, and the algorithm monitors the torque profile. When it detects a sudden increase — that's the screw head contacting the surface — it switches to seat mode, which ramps up torque to the final spec and then stops. Luna: And the 8 percent failure rate — what went wrong? Lucas: Mostly cross-threading on the first engagement, especially with smaller screws — sizes under M2. The robot's search pattern sometimes didn't align the screw perfectly perpendicular to the hole, so the threads would catch at an angle. Also, there were cases where the screw was slightly damaged or had burrs — the torque signature looked normal, but the screw wouldn't seat properly. Luna: So it's still not reliable enough for a production line where you might have tens of thousands of screws per shift. Lucas: Exactly. 92 percent sounds great in a lab, but in manufacturing, you need six-sigma reliability — 99.9997 percent. A single cross-threaded screw can scrap an entire assembly. So human oversight is still required. Luna: But the paper shows that the gap is closing. What was the hardware setup? Lucas: They used a standard six-axis industrial arm — a Fanuc LR Mate 200iD — with a custom gripper that held a commercial electric screwdriver. The force-torque sensor was an ATI Mini45, which is pretty standard in research labs. The algorithm ran on an external PC at 1 kHz loop rate. They also used a camera for initial screw detection, but not for the driving itself. Luna: So the sensing was all through the wrist, not vision. Lucas: Right. And that's actually the more robust approach for industrial settings, where lighting and debris can mess up vision. The torque signature is a richer signal — it tells you about the state of the screw joint in real time. Luna: I wonder how that compares to what companies like Tesla or Toyota are doing internally. They must have proprietary systems. Lucas: They do, but they're usually built around specialized end-effectors — like a nutrunner with a built-in torque sensor and a compliance mechanism. Those are expensive and specific to one fastener type. The MIT approach is more general-purpose: the same robot arm can switch between screw sizes without hardware changes, just by updating the algorithm parameters. Luna: That's the holy grail for flexible manufacturing — one robot that can do multiple assembly tasks without retooling. Lucas: Exactly. And that's where the research is heading. The next step, according to the authors, is to incorporate learning from demonstration — show the robot one human-performed screw driving, and let it generalize to new screw types. They're also working on detecting stripped threads before the screw is fully driven, which is a much harder problem. Luna: Stripped threads — you'd need to sense the torque dropping off unexpectedly, right? Lucas: Yes. When a thread strips, the torque peaks and then suddenly drops. But the challenge is distinguishing that from normal variations due to lubrication or surface finish. They're training a neural network on torque-time series data from thousands of screw driving events. Luna: So it's a data problem as much as a hardware problem. Lucas: Absolutely. And that's a theme across robotics right now — the hardware is getting better, but the software and data are where the breakthroughs are happening. Luna: I think it's also worth mentioning that screw driving is not just about the arm — the screw itself matters. There's a whole field of fastener design optimization for automated assembly. Lucas: Great point. Screws with pointed tips, for example, are easier to align. Thread-forming screws that cut their own threads require different torque profiles than machine screws. And coatings like dry lubricant can change the friction coefficient dramatically. So the robot has to be robust to those variations. Luna: Which brings us back to why humans are still better — we can adapt on the fly. If a screw feels tight, we back off and try again. A robot might just keep pushing and break something. Lucas: Yeah, and that adaptive behavior is hard to encode. But it's not impossible. There's work on using reinforcement learning to teach robots to 'wiggle' a stuck screw, similar to what a human would do. Luna: We should probably look at the broader picture — why does this matter beyond the lab? What industries would benefit most? Lucas: Consumer electronics is a big one. Smartphones, laptops, smartwatches — all have dozens of tiny screws. Currently, most of those are assembled in factories in China and Taiwan, with workers using electric screwdrivers. Automating that could bring some assembly back to other regions, because you don't need cheap labor if the robot can do it. Luna: Also medical devices — screws for orthopedic implants, surgical instruments. Those require extremely precise torque to avoid damaging the device. Lucas: Right. And aerospace, where every fastener has a specific torque spec and a human has to sign off on each one. If a robot could log the torque curve for every screw, that would be a huge audit trail advantage. Luna: So there's a clear ROI case — especially as labor costs rise and quality demands increase. Lucas: Exactly. And that's why companies like ABB and Kuka are investing heavily in force-controlled assembly. They see screw driving as a gateway capability — if you can master that, you can do many other precision assembly tasks. Luna: It's interesting that something so mundane is actually a frontier of robotics research. Lucas: It really is. And you know, if today's tech conversation gave you something usable, a handful of listeners already chip in monthly through buy me a coffee dot com slash fexingo. That's what keeps this show ad-free and lets us dig into these niche topics without any sponsor pressure. Luna: Yeah, it's a small group, but it makes a big difference. We appreciate every one of them. Lucas: So back to the research — one of the more surprising findings in that MIT paper was that the robot actually performed better with slightly worn screws than with brand new ones. The worn screws had a more consistent torque profile. Luna: That is surprising. I would have thought new screws would be more uniform. Lucas: Turns out, new screws often have small burrs from the manufacturing process that cause friction spikes. After a few uses, those burrs wear down, and the torque curve smooths out. So the robot's algorithm, which was trained on data from used screws, actually generalized better to worn ones. Luna: That's a great example of why real-world testing is so important. Lab conditions can be misleading. Lucas: Absolutely. And it's also a reminder that the problem isn't just the robot — it's the entire system, including the quality of the fasteners. If you want to automate screw driving, you need tighter tolerances on the screws themselves. Luna: Which adds cost upstream. But if you're assembling millions of devices, it might be worth it. Lucas: Yeah, it's a trade-off. And that's the kind of systems-level thinking that separates a lab demo from a factory deployment. Luna: So where do you see this going in the next, say, five years? Lucas: I think we'll see the first fully automated screw-driving cells in electronics assembly within three years, but they'll still have a human monitor overseeing multiple robots. The real breakthrough will come when the robot can detect and recover from failures autonomously — like if it cross-threads, it can back out and retry. That's probably five to seven years out. Luna: And by then, we might be saying 'remember when robots couldn't drive a screw?'