Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Pick a Single Screw from a Bin
Transcript
- Lucas: Let me paint a picture. You're in a factory that assembles automotive sensors. There's a bin — maybe a foot square, six inches deep — filled with M6 hex-head screws, washers, and nuts, all jumbled together. A human worker reaches in, finds the right screw by feel, and picks it out in under two seconds. A state of the art robot arm with a vacuum gripper and a depth camera might take five seconds and still drop or misidentify parts 30 percent of the time. Luna: Thirty percent failure? That's not even close to production-ready. Why is it so hard? Lucas: It's a problem called bin picking, and it's been on the robotics research agenda since the 1980s. The core issue is perception: the robot needs to see a 3D scene with overlapping, reflective, textureless objects and figure out exactly where one cleanly graspable surface is. The depth camera sees a mess of points, and the segmentation algorithms often can't tell where one screw ends and a washer begins. Luna: But we have deep learning now. Why can't we just train a neural net on a million images of screws in bins? Lucas: That is what people have tried. And it works—up to a point. A 2025 study from the Fraunhofer Institute for Production Technology tested six different deep-learning-based bin-picking systems on a set of 50 common industrial parts. The best system achieved an 82 percent success rate on parts it had seen during training. But on novel parts, that dropped to 58 percent. And the failures were often catastrophic—the robot would drive its gripper into the bin because it estimated the depth wrong. Luna: So the models don't generalise well. That's a data problem, isn't it? Lucas: Partly. But there's a deeper issue: specular reflections. Screws and other metal parts are shiny, so the infrared pattern from a structured-light depth camera reflects away. The sensor sees holes where there's surface. And even with a high-end lidar, you get noise around edges. The latest work from UC Berkeley's AUTOLab is using something they call 'multi-focal depth cameras' — they take images at two different focus settings and combine them to get better edge fidelity. They report a 12 percent improvement in grasp success on shiny parts. Luna: Twelve percent is meaningful but still not enough. What about gripper design? Is it just a vision problem? Lucas: No, grasping is equally tough. A vacuum gripper needs a seal against the part surface. If the screw is tilted, the seal leaks. And a traditional parallel-jaw gripper needs to know the exact orientation to avoid colliding with neighbouring parts. Researchers at the German Aerospace Center have been working on 'tactile bin picking' — equipping gripper fingers with high-resolution tactile sensors that can confirm a grasp before lifting. In a 2024 paper, they showed that combining vision with tactile feedback improved first-time grasp success from 71 to 89 percent, but it added almost a second per pick cycle. Luna: That extra second adds up. In a high-volume line, you're throughput-limited. So what's the practical workaround right now? Lucas: Most factories still use vibratory bowl feeders or custom fixture trays to present parts in a known orientation. Bin picking is only deployed when the parts are so varied that feeding isn't practical. Fanuc and ABB both sell bin-picking cells, but they require extensive tuning per part family. The ROI only makes sense if you're changing parts often. But there's a strong pull from e-commerce logistics—think of picking random items from a tote—which is driving investment. Luna: And that pull is pulling in a lot of listener interest, actually. If today's tech conversation gave you something usable—a clearer picture of why your warehouse robot keeps dropping washers—consider supporting the show. We keep it ad-free thanks to listeners who visit buy me a coffee dot com slash fexingo. It's a small way to keep these deep dives coming. Lucas: Yeah, and we really appreciate that — lets us spend the time reading those Fraunhofer studies. Anyway, that UC Berkeley multi-focal camera is promising, but it's still in the lab. What do you think, will we see a robust bin-picking product within five years? Luna: I think we'll see it earlier, but only for narrow domains. For the general case—any part, any bin—that's a decade away, maybe more. The combinatorial complexity of shapes and materials is just huge. Lucas: And that's the sobering reality. Robotics progresses in hard-won percentage points, not breakthroughs. One day, maybe an M6 screw is just a pick and place. For now, it's still a research thesis.