Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Open a Stuck Drawer
Transcript
- Lucas: So, a listener named Priya wrote in with a question that's been bugging her at work. She's in a packaging facility, and they've got a robot arm that's supposed to pull open a drawer on a machine to load components. Works fine when the drawer's new and smooth. But once it's been used a few thousand times, it gets a little sticky, and the robot just... fails. It either yanks too hard and jerks the whole machine, or it doesn't pull hard enough and stops. She asked, why is this so hard for robots? Luna: I love that question because it sounds so mundane, but it's actually a nightmare in robotics. I mean, we've talked about wet fishing lines, slippery fish, zip ties... this is like the sticky drawer episode. Lucas: Exactly. And the reason it's hard is all about friction and force control. When you or I open a slightly stuck drawer, we don't just pull with a constant force. We instinctively wiggle it, apply a bit of extra force, feel it give, and then ease off. We use the tactile feedback from our fingers and the resistance we feel to adjust in real time. Luna: Right, we're basically doing a mini closed-loop control system without even thinking about it. Lucas: Yeah, and that's exactly what a robot has to be programmed to do, but it's not trivial. Traditional industrial robots, especially the ones you see in factories, are what we call position-controlled. They're really good at moving to a precise spot with incredible repeatability. You tell it to go to X, Y, Z coordinates, and it does that. But they're stiff. They don't give. So when the drawer resists, the robot just keeps pushing or pulling to its commanded position, and either the motor stalls or it applies way too much force and the whole thing lurches. Luna: So it's not that the robot can't pull a drawer. It's that it can't pull a drawer that gives a little. Lucas: Precisely. And that's where force control comes in. Instead of just controlling position, you're controlling the force the robot applies. There's a technique called impedance control, which is kind of the gold standard for this. The idea is that you don't command a movement, you command a relationship between motion and force, sort of like a virtual spring. So the robot can comply with the environment, push until it meets resistance, and then adjust. Luna: Okay, so impedance control is like teaching the robot to feel its way. But how does that actually work in practice for a stuck drawer? Because there's a specific moment when the drawer breaks free, right? Lucas: That's the crux. The static friction is higher than the sliding friction. So you need to apply enough force to overcome the initial stickiness, but then once it moves, you need to instantly reduce the force so you don't yank the drawer out. Humans do this seamlessly. For a robot, it's a real control problem. You've got to be sampling the force feedback fast enough and adjusting the command in real time. Luna: And the failure mode you mentioned — where the robot stops because it thinks it's hit a wall — that's because it's programmed with a force limit. It's like a safety threshold. It's not the same as 'I feel resistance, let me try a little more.' Lucas: Right. Many industrial robots have a force limit to protect people and equipment. So if the drawer is a bit stuck, the robot hits that force limit, thinks something's wrong, and just stops. It doesn't have the nuance to say, 'This is a known friction, let me try a bit harder.' Luna: So what's the solution? Is there a robot arm that can actually do this reliably now? Lucas: There's been a lot of progress, especially in collaborative robots, or cobots, which are designed to work alongside humans. They've got torque sensors in each joint, which gives them that force feedback. But even with that hardware, the software has to be clever. Researchers are using machine learning to train robots on the task of opening drawers and doors. They collect a lot of data from different scenarios — different levels of stickiness, different orientations — and the robot learns a policy that maps the force feedback to an action. Luna: So instead of hard-coding a wiggle motion, the robot learns it from experience. That's kind of like how we learn to open stuck drawers. Lucas: Exactly. And there's a really interesting paper from a few years ago that used a technique called force-based reinforcement learning, where the robot tries different actions and gets a reward for successfully opening the drawer without damaging it. Over time, it learns the right strategy: apply a bit of force, feel for the release, and then pull smoothly. Luna: But I imagine there's a big gap between the lab and the factory floor. Because in the lab, you can control everything. In a real factory, that drawer might be sticky in a different way every time. Lucas: That's the million-dollar question. And it's why this is still an active research area. The lab tests are promising, but the robustness isn't there yet. The robot might handle a specific type of stickiness well, but if the drawer gets a little warped from temperature changes or a piece of debris gets in the track, it can still fail. Luna: So what's Priya supposed to do in the meantime? She's got a robot that's just stopping mid-task. Lucas: A pragmatic fix is to add a small compliance device, like a spring-loaded mechanism between the robot and the drawer handle. That gives you a bit of passive compliance, so the robot doesn't feel the full resistance. It's not as elegant as force control, but it can get you to 90 percent of the way. Another approach is to upgrade to a cobot that has force sensing built in, and then tune the impedance parameters to match the specific drawer. Luna: That's a good point. And I've got to say, this whole topic reminds me of something I was reading about the other day. It's from an old robotics textbook, but there's this concept of 'compliance' in robot arms that's been around for decades. The PUMA robot in the 80s had a remote center compliance device that was basically a mechanical spring to handle peg-in-hole tasks. So this isn't new, but the sensors have gotten so much better. Lucas: Right. And that's what makes this exciting. The hardware is finally catching up to the theory. We've got force-torque sensors that are cheap enough to put on a lot of robots, and the processing power to run the control loops fast enough. So the stuck drawer problem is becoming more tractable. Luna: And speaking of that, I know our listener base is full of engineers and tinkerers. If they're dealing with a similar problem, what's a good starting point to actually implement a force control loop on a hobby robot? Lucas: Well, you don't need a six-axis arm to start. You can get a simple linear actuator with a load cell and an Arduino. Write a basic impedance controller that reads the force and adjusts the velocity. Start with a drawer that's not too stuck, and then gradually increase the friction. It's a great way to learn the concept hands-on. Luna: That's a perfect project for a weekend. And actually, I've been thinking about this episode — we're always talking about the limits of robots, and it's easy to get frustrated. But this is a problem that's being solved step by step. It's not like flying cars where we're waiting for a grand breakthrough. It's just a hard engineering problem. Lucas: Yeah, and that's what I love about this field. It's not about magic, it's about incremental progress. Every time we figure out one of these seemingly trivial tasks, it opens up new applications. I mean, think about how many tasks involve pulling something that doesn't move smoothly — opening a car door, a dishwasher, a filing cabinet. If we can crack this, it's huge for automation. Luna: And I have a feeling we'll be talking about this again when they figure out the next hard thing. But before we get too deep into that, I want to pause for a second. We've been talking about the value of this kind of knowledge — the nitty-gritty of robot control. And it's exactly that kind of deep dive that keeps this show ad-free and accessible. If you've ever gotten something useful from an episode like this, and you want to see more, you can support us at buy me a coffee dot com slash fexingo. It's a small gesture, but it genuinely helps us keep the lights on and the conversations going. Lucas: Yeah, and it's not about perks or anything — it's just knowing that people appreciate what we're doing. We've had listeners say they've used our episodes to explain robotics to their kids, or to make a case to their boss for a new cobot. That's the stuff that makes it worth it. Luna: Exactly. So if you're listening and you've found value, consider that. And now, back to the stuck drawer. So we've talked about impedance control, learning from data. But there's another angle that's been on my mind: the role of vision. Doesn't the robot need to see the drawer to know where to grab it? Lucas: That's a good point. Most of this stuff assumes the robot knows where the drawer is. But in a real environment, the handle might be slightly different, or the drawer might not be perfectly aligned. So you need vision to locate the handle and estimate its pose. And then you need to combine that with force feedback once you make contact. Luna: So it's a multi-modal problem: vision and touch. And that's where it gets really hard, because if the vision is off by even a few millimeters, the force feedback will be different than expected. Lucas: Exactly. And there's also the question of what happens when the robot's gripper slips on the handle. That's a whole other problem. We've talked about slippery fish, but a smooth metal drawer handle is also tough. Some robots use a suction cup, but that doesn't work for every handle. Luna: I remember that episode on the wet fishing line. This is like the inverse — a dry, smooth handle. Lucas: Yeah, and it's interesting because the solution might be the same: better tactile sensing. If the robot can feel the slip before it loses grip, it can adjust. That's a whole area of research in tactile sensors. Luna: So, do you think in the next five years we'll see robots that can open any drawer, in any factory, without fail? Lucas: Honestly, I think we'll see it in controlled environments sooner than we think. But for arbitrary drawers in the wild? That's still a challenge. There are just so many variables — size, weight, friction, alignment. But the trend is clear: robots are getting better at feeling their way through the world. Luna: And that's the story we'll keep following. Thanks for tuning in, and if you have a question about a robot that just can't seem to do something, send it to us. Lucas: We'll be back next week with another deep dive. Until then, keep an eye on that drawer at home — maybe give it a little wiggle and appreciate what your hand does without thinking.