Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Stapler
Transcript
- Lucas: So I tried a little experiment this morning. I walked into the lab, grabbed a standard office stapler off the shelf, and handed it to one of our research arms. Luna: Stationary robot arm, gripper, standard two-jaw parallel gripper? Lucas: Exactly. A Fanuc LR Mate, with a pneumatic gripper. And I asked it to pick up the stapler and staple a piece of paper. Luna: And I'm guessing it didn't go well. Lucas: It went about as well as you'd expect from a robot that can't see the staple exit hole, can't feel the resistance of the paper stack, and has to coordinate a swing that's part rotation, part translation, part force application. It jammed on the first attempt. Luna: Before we dive into that — you know we don't run ads on this show. We keep it clean because we want the focus on the actual engineering. If you find these deep dives valuable, the simplest way to support that choice is buy me a coffee dot com slash fexingo. Just a small gesture that keeps us independent. Lucas: Exactly. It matters because these episodes take real time to research. And we'd rather spend that time on the nuance than on ad copy. So if you're able, that link is buy me a coffee dot com slash fexingo. Now back to the stapler. Luna: Okay, so the arm had the stapler in its gripper. What happened next? Lucas: It positioned the stapler over the paper — roughly aligned. Then it tried to apply downward force. But here's the thing: stapling isn't just a linear press. The top part of the stapler pivots on a hinge, and the mechanism compresses a spring that drives the staple through the paper and into the anvil. The robot has to push at the right angle, with the right speed, and maintain force long enough for the staple to pierce and clinch. Luna: So it's not just force control, it's timing and trajectory. Lucas: Right. The robot pushed too fast, the staple bent sideways. It tried again slower, but then it didn't generate enough impact force to punch through. It's a classic contact-rich manipulation problem. And it's one that the Dobb-E project at Sony has been studying. Luna: Dobb-E — that's the learning from demonstration system they developed for household tasks? Lucas: Yes. They trained a robot on hundreds of human demonstrations of everyday tasks — opening drawers, picking up objects, but also using tools. And staplers came up. Their findings: even with a compliant wrist and visual feedback, the success rate on stapling was under 40 percent. The main failure modes were misalignment of the staple exit with the anvil, and insufficient downward force due to the robot not understanding the hinge mechanism. Luna: So the robot treats the stapler as a rigid object, but it's actually a compliant mechanism. Lucas: Exactly. The top half moves relative to the bottom half during use. Most robot grasp planners model the object as a single rigid body. They don't account for internal degrees of freedom. So the robot might grip the top half too high, and when it pushes down, the top slides without transferring force to the staple driver. Luna: Or it grips the bottom half, and then the top half is free to wobble. Lucas: Right. And even if the grip is correct, the robot has to sense the moment of contact between the staple and the paper — which requires high-bandwidth force feedback. Most industrial arms have force-torque sensors at the wrist, but they sample at maybe one kilohertz. That's plenty for steady-state force control, but for an impulsive event like staple penetration, you need faster sensing or a compliant mechanical design that absorbs the shock. Luna: So what would a robot need to staple reliably? Lucas: One approach is to use a soft, underactuated hand like the MIT SoftHand. That hand can conform to the shape of the stapler, distributing force over a larger area and allowing some internal slip. In tests, the SoftHand managed a 70 percent success rate on stapling, because the passive compliance compensated for misalignment. Luna: But 70 percent is still not production-ready. Lucas: No. For a warehouse or office robot, you'd need 99.9 percent reliability. But the interesting thing is that stapling is actually a harder benchmark than many of the tasks we've talked about on this show. It combines tool use, force modulation, precision alignment, and dynamic impact. It's a better test of dexterity than, say, folding a towel, because it requires the robot to interact with a mechanism that changes state. Luna: So what's the state of the art today? Are any research groups close to solving this? Lucas: The most promising work I've seen comes from a group at Carnegie Mellon that uses reinforcement learning in simulation with domain randomization. They train a policy to staple by varying the friction coefficient, the staple stiffness, the paper thickness — all in simulation. Then they transfer to a real robot. Their success rate on a real stapler is around 85 percent now, but only for a specific stapler model. Luna: So we're still far from a general-purpose stapling robot. Lucas: Very far. And that's a good reminder of how much we take for granted. When I staple a document, I don't think about the angle, the force, the alignment. My hand and brain do it effortlessly. But for a robot, every one of those sub-tasks is an open research problem. Luna: And the complexity compounds when you consider different staplers. Some have ergonomic grips, some are heavy-duty, some are mini staplers. Lucas: Exactly. The shape, the spring tension, the anvil geometry — all vary. So a robot trained on one brand might fail on another. That's the curse of contact-rich manipulation. It's the same reason robots can't yet handle a wet bar of soap or a human hair. Luna: So maybe the humble stapler is the new benchmark for robotic dexterity. Lucas: I think it should be. It's ubiquitous, it's low-cost, and it forces the robot to deal with a real-world mechanism that isn't designed for robots. Plus, it's easy to set up in any lab. I'd love to see a standardized stapler challenge at ICRA or RSS. Luna: That would be a great way to compare different manipulation approaches. And it would push the field toward solving tool use, not just pick and place. Lucas: Exactly. Because picking up a stapler is one thing. Using it is another. And until robots can do both reliably, we're still in the early days of dexterous manipulation. Luna: Well, I'm going to go try to teach our lab arm to staple. Wish me luck.