Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Single Toothpick
Transcript
- Lucas: So we have done 79 episodes of Why Robot Arms Still Can't handle various things — wet towels, raw eggs, fitted sheets. But one of our listeners, a mechanical engineer named Priya from Bangalore, emailed us a challenge she actually faced on the job. Her team was trying to automate a micro-assembly line that involved picking up wooden toothpicks and placing them into pre-drilled holes. And the robot arm just could not do it reliably. Luna: A toothpick. It's literally a stick of wood. Why can't a robot arm pick that up? Lucas: Right, that is exactly the question. So we looked into the physics. A toothpick weighs about 0.1 grams. It's roughly 65 millimeters long and two millimeters in diameter. And the surface is smooth but not perfectly smooth — it has a certain roughness from the wood grain. When a standard parallel-jaw gripper tries to grasp it, a few things go wrong. The first is called 'stick-slip friction'. Luna: Stick-slip — that's when the gripper jaws momentarily stick to the object and then slip suddenly, right? Lucas: Exactly. With a large, heavy object — say a steel beam — the normal force from the gripper is high enough that the frictional force dominates, and any stick-slip is negligible relative to the object's inertia. But with a toothpick, the normal force is tiny — you cannot just squeeze harder because you will snap the toothpick. So the gripper has to apply a very delicate force, and at that scale, the surface roughness causes the gripper pads to alternately stick and slip, leading to unpredictable micro-movements. The toothpick can rotate or even shoot out of the gripper like a watermelon seed. Luna: And that rotation is catastrophic if you're trying to insert it into a hole with only 0.1 millimeter clearance. Lucas: Yes. So Priya's team tried different gripper materials — silicone, rubber, even a custom foam — but nothing solved the fundamental issue. Then there is the problem of 'proprioception'. An industrial robot arm knows where its end effector is based on joint encoders. But when you're grasping a toothpick, the compliance in the gripper pads and the tiny deflections in the toothpick itself mean the actual position of the toothpick tip relative to the robot's kinematic model can be off by several millimeters. That is a huge error at this scale. Luna: So the robot thinks it's holding the toothpick centered, but it's actually off by half a toothpick width. Lucas: Right. And this is not just an industrial problem. In 2025, researchers at MIT's Manipulation Lab published a study where they had a KUKA LBR iiwa — that's a lightweight collaborative robot — attempt to pick up a single wooden toothpick from a flat table 500 times. Success rate: under 4 percent. And when they did succeed, it was often because the toothpick happened to land in a specific orientation that aligned with the gripper's sweet spot. Luna: That is astonishingly low. So what's the solution? Do we need better grippers, better sensors, or a completely different approach? Lucas: Probably all three. But the most promising direction I've seen is 'tactile servo control' — using high-resolution tactile sensors on the gripper surface that feed back real-time force distribution data, not just binary contact. A team at the University of Washington demonstrated a system where the gripper could detect the exact orientation of a toothpick after initial contact and then adjust its grip angle dynamically before lifting. They got success rates up to about 60 percent. Luna: Sixty percent is still not production-ready, but it's a huge leap from 4 percent. What about using suction instead of gripping? Lucas: Suction was tried. The problem is that toothpicks are porous, so vacuum leaks. And the surface area is too small for a reliable seal. Some teams have tried using electrostatic grippers, but those require high voltage and can damage sensitive components. So for now, micro-assembly lines that handle things like toothpicks still rely on human operators. And that is expensive and slow. Luna: It really shows how far we are from true human-level dexterity. I mean, a five-year-old can pick up a toothpick without thinking. Lucas: Absolutely. And that's the humbling reality. Our robot arms can weld car bodies with millimeter precision, but they cannot perform a task that requires coordinated force control at the sub-Newton scale with uncertain geometry. And this has real consequences beyond toothpicks — think about surgical suturing, where you're handling needles that are similar in scale, or assembling miniature electronics, or even food handling where you need to pick up a single grain of rice. Luna: Grain of rice — we actually did an episode on that. But toothpicks add their own twist because of the anisotropy of wood grain. Lucas: Right. And the fact that it's a long, slender object means the moment of inertia is very different along different axes. So if the gripper applies a slightly off-center force, the toothpick will tend to rotate to align with the direction of least resistance. That is actually a known problem in 'peg-in-hole' insertion, but with a toothpick, the aspect ratio is extreme — 30 to 1 length to diameter — so the alignment challenge is magnified. Luna: I recall that in the classic peg-in-hole task, the chamfer on the hole helps guide the peg in. But with a toothpick, the tip is blunt, and the holes Priya was using had no chamfer. Lucas: Exactly. So you have a blunt tip, no chamfer, and the robot's positional uncertainty is larger than the clearance. It's a perfect storm. And this is why I think the toothpick problem is actually a really good benchmark for the robotics community. It's cheap, it's standardized, and it exposes weaknesses in force control, tactile sensing, and insertion strategies. Luna: It is also a great test for learning-based approaches. Have any deep reinforcement learning methods been applied to toothpick grasping? Lucas: Yes, a group at Stanford trained a policy in simulation using domain randomization — they varied friction coefficients, toothpick lengths, gripper pad stiffness — and then transferred it to a real robot. Their success rate was around 35 percent, which is better than classical control but still far from reliable. The sim to real gap is particularly large here because the contact physics at this scale is hard to model accurately. Things like van der Waals forces and electrostatic attraction start to matter when objects are that light. Luna: So even the simulation doesn't capture the full reality. Lucas: Right. And that is a big research challenge. But I think the toothpick problem is getting more attention now because of its implications for automation in high-value industries like medical device assembly. For example, assembling a stent involves handling tiny wires that are similar in scale to a toothpick. If we can solve toothpick grasping, we can likely solve those problems too. Luna: And in the meantime, human hands remain the most versatile grippers ever invented. Lucas: Which is a nice segue into something I wanted to mention. If today's tech conversation gave you something usable — even just a better understanding of why your robot vacuum still bumps into chair legs — it's because we have the freedom to pick exactly these kinds of niche rabbit holes. And that freedom comes from a small group of listeners who chip in through buy me a coffee dot com slash fexingo. It's not a big ask — it's just what keeps the show ad-free and lets us spend a whole episode on toothpick physics. Luna: I think that is one of the coolest things about this community. People who care about the details, and who want to hear the real engineering behind headlines. It makes a difference. Lucas: Yeah. So, back to the toothpick. One of the more creative solutions I saw came from a startup in Zurich — they used a two-stage approach: first, a coarse positioning with a standard gripper, then a fine alignment using a micro-manipulator that literally vibrates the toothpick at ultrasonic frequencies to reduce friction during insertion. Think of it like an ultrasonic toothbrush, but for a toothpick. Luna: That is clever. Does it work? Lucas: In the lab, yes — they achieved about 80 percent success. But the ultrasonic vibration adds complexity and cost, and it can damage the toothpick or the hole if the vibration couples in the wrong way. So it is not a silver bullet. Luna: It sounds like there is no single solution. It's a system-level challenge — gripper, sensing, control, and even the design of the toothpick itself. Lucas: Exactly. And that is why I find this problem so fascinating. It forces you to think across the entire robotics stack. And it also reminds us that for all the progress in AI and autonomous systems, the physical world is still full of surprises. A two-dollar bag of toothpicks can humble a hundred thousand dollar robot. Luna: Priya from Bangalore — if you're listening, I think we owe you a coffee. This was a great topic. Lucas: Absolutely. And for everyone else — next time you pick up a toothpick, maybe give a little nod to how hard that actually is. Thanks for listening.