Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Can't Feel a Knife Blade
Transcript
- Lucas: So Luna, quick question. When was the last time you sharpened a kitchen knife? Luna: Honestly, I think I've done it twice in my life. And both times I was terrified I'd ruin the edge. Lucas: Right. And that fear is exactly the problem for a robot. Because sharpening a knife isn't just about applying force, it's about maintaining a precise angle while feeling the edge. And that's a sensory challenge that most industrial robots just aren't built for. Luna: So today we're going after the knife. Not the potato, not the screwdriver. The blade itself. Lucas: Exactly. And here's the hook. A standard force-torque sensor in a robot wrist can measure forces down to about 0.1 newtons. That's roughly the weight of a paperclip. But when you're sharpening a blade, the forces at the edge are tiny, and they change faster than a conventional sensor can sample. Luna: So the robot is essentially blind. It's pushing a blade against a whetstone without knowing if it's going to catch or slide. Lucas: And the worst part is that a knife edge is the perfect worst-case scenario. It's a line of contact, not a surface. So all that force is concentrated on a razor-thin strip. If the robot pushes a little too hard, it damages the edge. If it goes too soft, it does nothing. Luna: I remember you mentioned in an earlier episode that a robot can't even feel a screwdriver. How is a knife any different? Lucas: Good question. With a screwdriver, the main issue is sensing torque and slip. With a knife, it's about detecting the edge's geometry. You need to know the angle of the bevel, the burr formation, and the exact point of contact. And that's not just force, it's spatial resolution. Luna: So we're talking about tactile sensing, not just force sensing. Lucas: Precisely. And the state of the art in tactile sensing has improved a lot. There are arrays that can detect pressure changes down to a few millipascals, with spatial resolution under a millimeter. But they're still not good enough to feel a burr on a blade that's just a few microns wide. Luna: Microns. So we're not even in the millimeter range. That's like feeling a single grain of sand on a table. Lucas: And the challenge is that the sensor has to be robust enough to survive contact with a knife blade. Most tactile skins are made of soft silicone, which tears easily. So you have this trade-off between sensitivity and durability. Luna: So what are the actual approaches people are taking to solve this? I imagine it's not just about better sensors. Lucas: One promising line of work is using force-torque sensors at the wrist combined with machine learning. Instead of trying to feel the edge directly, the robot learns from the vibration and the sound of the blade on the stone. It's like listening for the pitch of the metal singing. Luna: That's clever. But there's still the problem of the blade moving. A knife isn't rigidly fixed in a robot's hand. It can flex, twist, and shift. Lucas: Right. So another approach is to add compliance to the gripper. Instead of a rigid metal hand, you use a softer, more flexible finger that can deform around the blade. That way, the robot doesn't need to know the exact position; it can just let the finger conform to the blade's shape. Luna: So instead of fighting the uncertainty, you embrace it. Lucas: Exactly. There's a research group in Japan that built a gripper with soft bellows fingers that can hold a knife at a consistent angle while rubbing it against a stone. They got decent results, but it's still far from human-level sharpening. Luna: And what about the company that makes those very dexterous hands, Shadow Robot? I've heard they have a tactile skin that can sense shear forces. Lucas: Shadow's skin is impressive. It uses a piezo-resistive material that can detect both normal and shear forces. But even that has a resolution of about one millimeter, which isn't enough to feel a burr. And it's still fragile. They've had to reinforce the fingertips with a metal shield just for basic handling. Luna: So we're hitting a wall. The sensors aren't sensitive enough, and the materials aren't durable enough. Lucas: And that's why this is such a rich area of research. It's forcing people to rethink the entire approach to robotic manipulation. Instead of trying to replicate human skin, some labs are exploring magnetic sensors or optical sensors that can measure deformation from the inside out. Luna: Interesting. Like embedding a camera inside a soft finger and watching the dots deform. Lucas: Exactly. That's the GelSight approach from MIT. They use a soft gel with a reflective coating and a camera underneath. When the finger presses on something, the gel deforms, and the camera sees the shape of the contact. That gives you both spatial resolution and force information. Luna: And that's the kind of thing that could eventually help a robot sharpen a knife. But we're still a long way from putting that in a commercial robot. Lucas: A long way, but not impossible. And you know, this whole episode made me think about the fact that we do this show for free, no ads, no sponsors. It's just us and the topic. And if you're getting something out of it, you can support the show at buy me a coffee dot com slash fexingo. It's a small thing, but it keeps us going. Luna: Yeah, we're really lucky to have listeners who appreciate that. And speaking of appreciation, I appreciate that you didn't say 'we're not a robot hand' when I asked about the knife. Lucas: Ha. Anyway, back to the blade. So there's also the question of safety. In a kitchen, a knife is dangerous. You can't just have a robot swinging a knife around without a way to detect contact with a human. That's where force sensing becomes critical. Luna: So even if you solve the tactile problem, you still need the robot to stop or retract if it touches something unexpected. Lucas: And that's why collaboration robots, or cobots, are designed with current limiting and torque sensing. They're inherently safer, but they're also less precise. So you have a trade-off: safety versus rigidity. Luna: So what would the ideal robot knife handler look like? Maybe a soft gripper with high-resolution tactile skin, a fast vision system to track the blade, and a compliant wrist. Lucas: That's the vision. And there are prototypes that combine all of those. There's a startup in Switzerland that's working on a robotic chef that can chop vegetables with a real knife. They use a custom force-torque sensor at the wrist and a machine learning model that predicts the blade's motion. Luna: But they're probably not sharpening the knives themselves yet. Lucas: No, they send them out for sharpening. But they're working on it. The point is that the gap between human and robot dexterity is closing, but the last mile is the hardest. Luna: So for now, the knife remains one of the ultimate challenges for robotic hands. It's a perfect test of sensitivity, durability, and intelligence. Lucas: And that's why we'll probably be talking about robot hands and knives for a while. But next time you pick up a chef's knife, just appreciate the incredibly complex feedback loop between your fingers, your eyes, and your brain. Luna: I will. And I might also appreciate that I don't have to program a robot to do it.