Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Sort a Bin of Mixed Parts
Transcript
- Lucas: So you know that classic robot demo where the arm reaches into a bin and pulls out a specific part, over and over, like it's nothing? Luna: Yeah, the ones that make it look effortless. I've seen the videos. Lucas: Right, well, in a real factory, that bin is rarely full of identical, neatly arranged screws. It's a chaotic jumble of different sizes, shapes, and materials — and that's the problem that's been haunting robotics for three decades. Luna: So today we're asking: why can't a robot arm just sort a bin of mixed parts? Lucas: Exactly. And the short answer is that it's not about vision — it's about touch. You can point a high-res camera at a pile of bolts, but the robot still doesn't know how hard to squeeze, or where the part will shift when the gripper closes. That's where everything falls apart. Luna: So it's a physics problem, not just a perception problem. Lucas: Precisely. Let's break it down. The classic approach is what's called 'random bin picking.' The robot uses a 3D camera to identify a part's pose, plans a grasp, and goes for it. But in a mixed bin, the parts occlude each other, they're tangled, and sometimes a grasp that looks perfect in simulation just crushes the part or slips right off. Luna: And that's where force sensing comes in? I've heard that term thrown around a lot. Lucas: It's the missing ingredient. A force-torque sensor at the wrist measures the actual contact forces in real time. That lets the robot feel when a screw is jammed or when the gripper is about to slip. And that feedback loop is what's finally moving the needle. Luna: So what changed recently? I'm guessing it's not just better sensors. Lucas: Right, it's the algorithms that use those sensors. Last quarter, a German automation firm published a bin-picking system that cleared a real factory's fastener bin — mixed M6 and M8 bolts, plus some washers — at a 99.1 percent success rate. That's a huge jump from the typical 85 to 90 percent we've seen before. Luna: Wow, 99.1 percent. What's their secret? Lucas: They trained a neural network in simulation — literally millions of attempts — but then they fine-tuned it on the real robot's force feedback. The system learns to recognize when a grasp is going to fail before it even applies full force, and then it adjusts the approach angle or pressure on the fly. Luna: So it's not just pattern recognition, it's learning from its own mistakes in the real world. Lucas: Exactly. And the key insight was that they stopped trying to plan the perfect grasp ahead of time. Instead, the robot does a quick exploratory touch — a little tap or wiggle — to feel out the part before committing. That's something a human does instinctively, but it's completely counterintuitive to how most robot control is written. Luna: Okay, so why hasn't anyone done this before? It sounds obvious now. Lucas: The obvious is always the hardest. For decades, the industry focused on improving vision systems — better cameras, faster point clouds. But vision can't tell you if a part is stuck under another one. Force sensing is the missing sense, and it's only in the last few years that affordable sensors and the compute power to process that data in real time have come together. Luna: And that's the thing that makes this a business story, not just a tech story. The cost of these sensors has dropped dramatically — we're talking below two thousand dollars for a decent six-axis unit now, versus ten to fifteen thousand a decade ago. Lucas: Right, and that's opening up the market. It's not just automotive giants anymore. Mid-sized manufacturers can now afford a bin-picking cell that pays for itself in a couple of years, especially in areas where labor is hard to find. Luna: Let's dig into the specifics of that German system — what was the actual setup? Did they use a special gripper or just a standard one? Lucas: Good question. They used a standard two-finger parallel gripper with a force-torque sensor at the wrist. Nothing exotic. The magic was in the control software, which runs at about a kilohertz — that's a thousand decisions per second, adjusting the grip based on tactile feedback. They also used a compliance strategy: if the part doesn't move as expected, the robot backs off and tries a different angle. Luna: So it's about being gentle and adaptive, not just strong and fast. Lucas: Exactly. And that's a shift from the traditional 'force the issue' approach that dominated industrial automation. The system also has a recovery routine: if it fails to pick after three attempts, it explicitly searches for a new part by nudging the pile, not by blindly plunging in. Luna: That sounds like it's learning, but is it actually learning on the job, or is it pre-programmed with those recovery behaviors? Lucas: There's a bit of both. The exploration behavior is learned in simulation — the system discovers that nudging helps untangle parts. But the real-time adaptation to slip or jam is reactive, based on force thresholds. And importantly, they don't require the robot to generalize to completely new parts without retraining. It's a narrow application, but it's robust within that narrowness. Luna: So it's not a general-purpose solution yet. But for a factory that only handles a few types of fasteners, that's fine. Lucas: Exactly. And that's the realistic path for robotics right now — not some universal humanoid, but highly capable systems that do one job really well. The economic payoff is immediate: that firm's customer reported a 30 percent reduction in manual handling injuries in the first six months, because workers no longer had to reach into bins to untangle parts. Luna: That's a compelling ROI story — safety, efficiency, and cost all at once. Lucas: So, if today's tech conversation gave you something usable — a clearer picture of why bin picking is so hard, or a sense of where automation is actually heading — the way we keep this show ad-free is through listener support. If that's valuable to you, consider buying us a coffee at buy me a coffee dot com slash fexingo. Luna: Yeah, that's the kind of support that keeps us digging into these stories, and it's just a small gesture that goes a long way. Lucas: Now, back to the technology. I want to talk about a surprising side effect of this success: it's forcing a rethink of how we design robot workspaces. Instead of assuming the robot will handle a perfectly organized bin, engineers are now designing bins and part presentations that are deliberately messy, because the robot can handle it — and it's cheaper than installing expensive feeders. Luna: So the robot is adapting to the mess, rather than the mess being engineered away. Lucas: Exactly. And that's a big deal for small and medium manufacturers, because vibrating bowl feeders and custom part orienters can cost tens of thousands of dollars and take weeks to change over for a new part. A bin-picking cell with the right software can be reconfigured in an afternoon. Luna: That's the real practical benefit — flexibility. But I'm curious, what are the limits? What's the next bin-picking challenge that's still unsolved? Lucas: The big one is handling deformable parts — think rubber gaskets, cables, or fabric. Those don't have a rigid shape that can be modeled easily. You can't just simulate a thousand attempts because the part's behavior is unpredictable. And force sensing helps, but it's not enough when the part can bend and twist in ways that change the contact model entirely. Luna: So we're back to why robots can't handle a wiring harness, which we did in episode 132. Lucas: Ha, yeah, we did. But that's the frontier. And there's also the problem of very small parts, like surface-mount electronics components. They're so light that the force sensor can barely register a difference between touching them and not touching them. That's a different kind of challenge. Luna: So the technology is getting better, but there are still clear boundaries. I think that's a realistic and helpful takeaway for our listeners. Lucas: Definitely. And that's the thing — every time we see a breakthrough like that 99.1 percent result, it's worth asking what the underlying principle is. Here, it's about using touch to compensate for imperfect vision. That's a lesson that applies beyond bin picking, to any task where the robot needs to interact with the physical world in a nuanced way. Luna: So next time you watch a robot struggle with a bin of mixed parts, you'll know it's not about the camera — it's about the sense of touch. Lucas: Exactly. And maybe, in a few years, it won't struggle at all.