Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Can't Feel a Screwdriver
Transcript
- Lucas: So we've spent the last twenty episodes talking about robot hands failing at everything from flipping pancakes to threading needles. But today, for our 150th episode, I want to drill into something that might be even more fundamental. The reason a robot arm can't reliably screw in a lightbulb -- we did that one -- but also why it can't do something as simple as turning a screwdriver in a tight space. The problem is that the robot can't feel the twist. Luna: Feel the twist. You mean torsion, right? The rotational force that tells your hand how the screw is seating or if it's stripping? Lucas: Exactly. And that's a really useful thing to talk about today, because it's the kind of problem that sounds like it should be solved -- we've had torque sensors for decades -- but it's still one of the biggest bottlenecks in automation. So, before we get into the weeds, a quick note. This podcast is ad-free and listener-supported, which is why we can spend fifteen minutes on something as niche as torsional compliance in robot wrists. If today's conversation gives you something useful, and you feel like throwing a coffee our way, the link is buy me a coffee dot com slash fexingo. It's the smallest possible ask, and it genuinely keeps the lights on. Luna: Yeah, and it's not like we're trying to sell you a robot arm. It's just a nice way to say thanks if you can. And honestly, a coffee is probably cheaper than the torque sensor we're about to talk about. Lucas: Ha, true. Okay, so let's get into it. The core problem is that most robot arms use a force-torque sensor at the wrist, and that sensor measures forces in six axes. But the sensor is usually a stiff, metal cylinder between the robot flange and the tool. It's rigid. And that rigidity is exactly why the robot can't feel a screwdriver twist. Luna: Because the sensor only sees the torque at the wrist, not the tiny variations at the tip. The screwdriver is like a long lever, and the sensor is sitting at the wrong end. Lucas: Right. And that's the first issue. But even if you put a torque sensor directly on the screwdriver bit, you have a bigger problem: the sensor itself is stiff, and stiffness kills the feedback. In a human hand, you have this beautiful combination of soft skin, flexible joints, and hundreds of nerve endings in your fingertips. When you feel a screwdriver twist, you're feeling a distributed deformation of your skin, not a point measurement. Luna: So you're saying the sensor gives you a single number, but your hand gives you a whole surface of data. And that surface data is what lets you adjust the force and angle in real time, without thinking about it. Lucas: Exactly. And that's why we've seen these research projects where robots use a standard force-torque sensor to try to screw in a screw, and they fail maybe twenty percent of the time. They can't detect the moment when the screw head seats against the material, or when the threads catch slightly. They just see a gradual increase in torque, and then a sharp spike, and by the time they react, the screw has already stripped. Luna: But don't some industrial robots do this successfully? Like, in car assembly lines, they've been using screwdrivers for a long time. Lucas: They do, but the key is that they add a compliant element. You'll see a robotic screwdriver that has a spring-loaded bit, or a little bit of give in the tool. That compliance absorbs the initial misalignment and gives the sensor enough time to read the torque. But it's not true feedback, it's a fudge factor. And it's fragile. The spring wears out, the calibration drifts, and you have to tune it for every single screw type and material. Luna: So it's a workaround, not a solution. And that's why a robot can't just pick up any screwdriver and use it like a human would. Lucas: Right. Let me give you a concrete example. There's a small automotive supplier in Ohio that makes door handles. They wanted to automate the final assembly, which involves inserting four small screws into a plastic housing. They hired a systems integrator, spent two years, and a few million dollars, and the line still has a manual station at the end. The robot can do about ninety-five percent of the screws, but the last five percent have tolerances that are too tight. The plastic holes are a little bit out of spec, and the robot can't feel the subtle difference between a screw that's going in straight and one that's about to cross-thread. Luna: And a human worker can just feel the resistance change and back off. The robot can't, because the sensor is too stiff and too slow. Lucas: Exactly. And that five percent is the difference between a fully automated line and one that still needs a person. That's the business case we keep seeing over and over. The robots are fast and precise, but they lack the sensitivity to handle variation. And that variation is what the real world throws at you. Luna: So what's the fix? Are there new sensors that are actually soft and fast enough to give real feedback? Lucas: There are a few promising directions. One is vision-based tactile sensors, like the ones that use a camera inside a soft gel dome. The camera tracks the deformation of the gel as it presses against an object. That gives you a high-resolution map of the contact surface, including slip and torsion. These are still mostly in research labs, but they're getting cheaper and faster. Luna: But there's a catch, right? You need to process all those images quickly enough to react in real time. And the gel gets scratched or dirty. Lucas: Yeah, that's the tradeoff. The sensor itself is clever, but you need a lot of compute to extract the torque estimates from the images, and the gel is a consumable. It's not ready for a factory floor yet. But the idea is that you can get thousands of tactile pixels, instead of one strain gauge reading. Luna: And then there's the deep learning angle. Instead of hardcoding the control logic, you train a neural network to map the tactile images to the right motor commands. That's what some labs are doing, and it's actually showing some promise. Lucas: Yes, and that's where the field is heading. But the training data problem is huge. You need thousands of hours of human demonstrations, or you need a simulation that accurately models the physics of the gel deformation. And that's still not easy. Luna: But I've also heard about using the robot's own motor currents as a proxy for torque. Is that not enough? Lucas: It's a start, but the signal is noisy and laggy. The motor current tells you the average torque over a few hundred milliseconds, but it's smoothed out by the inertia of the arm and the gearbox. So you can't detect the subtle sign of a thread catching. It's like trying to feel the texture of a surface through a thick work glove. Luna: That's a great analogy. And that's why the human hand is still so hard to beat. It's not just the sensors, it's the whole package of skin, muscle, and brain that processes it all in milliseconds. Lucas: And that package is also the reason why those earlier episodes -- the ones about peeling potatoes and folding towels -- all share the same root cause. It's not that the robot hand is dumb, it's that it's deaf and blind to touch. And touch is precisely what those tasks need. Luna: So for the 150th episode, we're basically saying that the bottleneck is the sense of touch. And that's a great way to sum up the last twenty episodes. Lucas: Right. And I think it's worth noting that there are some new startups trying to solve this. There's one from MIT that's building a soft, liquid-metal-based sensor that can be printed into a glove. It's still early, but the idea is to make tactile sensing cheap and scalable. Luna: That sounds promising. But how far away are we from a robot that can actually change a tire or assemble a phone without any special tooling? Lucas: Honestly, I'd say we're still five to ten years away from a general-purpose robot hand that can handle that kind of variability. But in the next two or three years, we'll see more specialized solutions that combine a better sensor with a compliant actuator, and they'll handle a narrow class of tasks, like inserting connectors or handling flexible cables. Luna: So it's not a silver bullet, but it's a definite improvement. And for companies that have a lot of these repetitive assembly jobs, even a small improvement in sensitivity could tip the ROI. Lucas: Exactly. And that's the business angle we like to dig into. It's not about building a robot that can do everything, it's about finding the one task where a twenty percent improvement in success rate saves you a hundred thousand dollars a year. Luna: And that's why we keep coming back to these niche tasks. Because they're the ones that actually matter to the people making things. Lucas: Yeah. So if you're walking away with one thing today, it's that the humble screwdriver is a deceptive challenge. It's not about brute force, it's about finesse. And finesse is hard to engineer. Luna: But not impossible. And I guess that's why we'll keep doing episodes like this. To see the moment when the sensors finally get good enough. Lucas: And when that happens, we'll be there to talk about it. Until then, I'm Lucas. Luna: And I'm Luna. Thanks for sticking with us for 150 episodes. Lucas: And remember, if you want to support the show, the link is buy me a coffee dot com slash fexingo. But no pressure. We'll be back next time with another robot that still can't do something you do without thinking. Luna: See you then.