Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Robot Hands That Can Feel a Screwdriver
Transcript
- Lucas: So we've spent fifteen episodes telling you what robot hands can't do. Peeling potatoes, snapping green beans, threading a needle. Today I want to talk about the one sense that might change most of those failures. Touch. Luna: Touch. As in, actual tactile sensing on a robot gripper? Because for a long time, most industrial arms just used position and force feedback, right? Lucas: Exactly. The classic approach is to control the joint torques and hope the fingers don't crush anything. But that's like trying to pick up a raw egg while wearing oven mitts, you can feel the pressure but you have no idea about the surface texture or whether it's slipping. Luna: So what's changed? I remember a few years ago, tactile sensors were laboratory curiosities. Now I'm hearing about them in warehouse robots. Lucas: Right. The big shift is capacitive tactile arrays. These are thin, flexible sheets with tiny electrodes that measure changes in capacitance when they deform. The output is a high-resolution pressure map, like a fingerprint of contact. Some arrays have hundreds of taxels, tactile pixels, per square centimeter. Luna: Taxels, okay. So instead of just knowing 'I'm touching something,' the robot gets a detailed picture of where and how hard it's pressing. That sounds like a game changer for something like picking a screw from a bin. Lucas: Yeah, that was one of our earlier episodes. The challenge there is not just feeling the screw but distinguishing it from the other screws and the bin itself. With a tactile map, the robot can recognize the shape of a single screw head, even if it's lying at an odd angle. Luna: But wait, didn't we also say that robots can't feel a screwdriver? I think that was episode 150. How is tactile sensing going to crack that? Lucas: Good memory. The screwdriver problem is about torque and slip. When you twist a screwdriver, the handle rotates in your palm unless you adjust your grip. A robot with only force control will either lose the handle or over-tighten and strip the screw. Tactile sensors can detect the micro-slip at the contact surface and adjust grip in milliseconds. Luna: So it's not just about knowing where the fingers are, it's about sensing the shear forces. That's the difference between holding a screwdriver and using it. Lucas: Precisely. And the technology is moving from research to real products. For instance, one startup I've been following makes a tactile sleeve that wraps around any gripper finger. It streams data at a kilohertz, which is fast enough to catch the vibrations of a screwdriver handle starting to rotate. Luna: A kilohertz. That's a thousand samples per second. Is that enough to feel a screwdriver twisting? I mean, humans feel texture at much higher frequencies, but for a tool handle, maybe that's plenty. Lucas: For a screwdriver, yes. The key is not just sampling speed but the algorithm that interprets the data. That's where machine learning comes in. You train a neural network on thousands of trials, teaching it to recognize the pattern of a successful grip versus a slip. Once it learns that, it can generalize to different handle widths and materials. Luna: So the sensor gives you the raw signal, but the intelligence is in the software. That's a interesting split from the old days where you tried to hard-code every possible scenario. Lucas: Exactly. And that's why tactile sensing is finally practical. The hardware is cheap enough, but the software is what makes it work. There's a famous demonstration from a lab at MIT where a robot with tactile fingers successfully inserted a USB cable into a port. That sounds trivial, but anyone who's fumbled with a USB in the dark knows it's a real challenge. Luna: I've done that more times than I want to admit. But how did the robot do it? Just feeling the shape of the plug? Lucas: Partly. The robot used tactile feedback to feel the contour of the port and the plug, and then it adjusted its insertion angle based on the contact forces. It's a bit like how you feel your way when you're trying to plug in a charger behind your desk. You can't see it, but your fingers guide you. Luna: So tactile sensing turns robots from blind machines into something more like a person fumbling in the dark. That feels like a step change, but are there any limitations? I mean, we've spent 151 episodes on what robots can't do. Lucas: Plenty of limitations. The tactile array I mentioned has a spatial resolution of about one millimeter, which is nowhere near human skin. And it struggles with very soft or very slippery objects, like a raw chicken breast or a wet towel. But for structured tasks in factories, it's already good enough. Luna: Talking about factories, I've seen videos of robot hands sorting mixed parts in a bin, and they're using tactile sensing to pick out specific items. That's something we said was impossible a few months ago. Lucas: Right. The bin-picking problem is a classic. With a tactile map, the robot can identify the orientation of a bolt or a connector, and then plan a grasp that won't slip. It's not perfect, but it's dramatically better than a few years ago. Luna: One thing I'm curious about: how do you test these sensors? I mean, you can't just put a robot hand in a lab and say 'feel this screwdriver.' What's the experimental setup like? Lucas: It's pretty rigorous. Researchers use calibrated objects with known surface textures and compliance. They'll run hundreds of trials, measuring the success rate of picking up, rotating, and placing objects. They also test for durability, because sensors that wear out after a thousand cycles aren't useful. Luna: And speaking of useful, I know we're an ad-free show, and that's something I really appreciate. We keep it that way because our listeners support us directly. If you want to help keep this podcast independent, there's a link to buy me a coffee at buy me a coffee dot com slash fexingo. It's a small gesture that means a lot. Lucas: Yeah, that's right. And we're grateful for every listener who chips in. It lets us keep doing deep dives without any corporate sponsor telling us what to cover. So thanks for considering that. Anyway, back to the tactile story. I was reading about a recent paper where they integrated tactile sensors into a prosthetic hand, and the user could feel the difference between a baseball and a tennis ball just by squeezing. Luna: That's wild. And that kind of feedback loop, where the sensor data is sent to the user's nerves, is a whole other frontier. It's not just about the robot's hand, but about restoring touch to humans. Lucas: Right. But for now, the focus in industry is on making the sensors more robust and cheaper. The current generation costs in the hundreds of dollars per finger, which is still a lot for a factory that needs dozens of grippers. But the prices are falling fast, similar to how LiDAR dropped over the past decade. Luna: So in five years, might we see every robot arm in a factory equipped with tactile sensing? Is that the trajectory? Lucas: I think so, at least for jobs that involve any kind of grasping or assembly. The market for tactile sensors is growing at a compound annual growth rate of around fifteen percent, driven by e-commerce automation and the push for more flexible manufacturing. We're not there yet, but the pieces are in place. Luna: It's exciting to think that some of those tasks we said were impossible might become routine. I'm still curious about the screwdriver, though. Can a robot with tactile sensing actually tighten a screw without stripping it? Lucas: There's a demonstration from a research group in Germany where a robot used tactile feedback to drive a screw into a block of wood at a controlled speed and torque. It detected the change in resistance as the screw bit into the wood and adjusted its force accordingly. The result was a clean, flush screw. No stripping. Luna: That's impressive. But what about a screw that's already stripped? Can the robot adapt? Lucas: That's a tougher one. A stripped screw has a non-uniform surface, and the tactile sensor would need to detect the irregularity and change strategy, maybe with a different tool or a bit of force. That's still a research challenge, not a commercial product. Luna: So we've covered the upsides, but I want to know the dirty secret. What's the catch? Why isn't every robot using this now? Lucas: The catch is durability and integration. The sensors are still delicate compared to steel grippers. They can tear or degrade with heavy use. And integrating them into an existing robotic arm requires changes to the control software, which is a significant engineering effort. Plus, there's the machine learning training time, which can be days for a new task. Luna: So it's not just plug and play. That explains why adoption is slower than you'd think. Lucas: Exactly. But the trend is clear. As the cost drops and the software improves, tactile sensing will become standard for any robot that handles unstructured objects. The question is not if, but when. Luna: And for someone listening who wants to see this in action, what's a good way to get a sense of the current state? Any public demos or open-source projects? Lucas: There are a few. The University of California has released open-source tactile sensor designs, and some companies sell evaluation kits. Also, check out robotic grasping benchmarks like the yale cmu berkeley dataset, which includes object sets and performance metrics. It's a great way to see what's actually working. Luna: I'll have to look into that. I love the idea of being able to test this technology myself, even if just in simulation. Lucas: And that's the beauty of this field right now. Between open-source hardware and simulation tools, a motivated hobbyist can get started with tactile sensing for a few hundred dollars. That wasn't possible five years ago. Luna: It feels like we're at an inflection point. Next thing you know, robots will be threading needles. But we'll save that for another episode. Lucas: Absolutely. For now, I'm curious to see how tactile sensing evolves in the next year. I'd bet that a robot will finally peel a potato without mashing it. That would be a first.