Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Warehouse Robots Fail at Picking Boxes
Transcript
- Lucas: There is a specific kind of frustration that only happens in logistics centers. You walk into a massive fulfillment warehouse, and you see rows upon rows of shelves perfectly stocked, robots gliding silently along tracks, packages flying through scanners. It looks like the future of work. But then you get to the packing station, or worse, the raw material bin, and everything stops. Luna: Because humans can still do it faster? Or because the robots just keep dropping things? Lucas: It is not just dropping things. It is the fundamental inability to pick a single object out of a jumbled pile without knocking over everything else around it. We have spent thirty years perfecting the assembly line, where parts come in trays with precise coordinates. But real life, and especially e-commerce returns or bulk shipping, is messy. It is a chaotic heap of irregular shapes, colors, and textures. This is what engineers call the bin picking problem, and it is currently the biggest bottleneck in scaling automated warehouses. Luna: So the tray method worked for cars but fails when you need to grab a random t-shirt from a mountain of other shirts. Lucas: Exactly. Think about Amazon or any major distributor. They used to rely heavily on human sorters for these unstructured tasks. But labor shortages and wage inflation are pushing them to automate. The challenge is physics. If a robotic arm tries to grab a box from the bottom of a stack, it has to calculate friction, weight distribution, and the structural integrity of the pile above it. One miscalculation, and the whole pile collapses, jamming the machinery or damaging the goods. Luna: And traditional cameras just see a wall of pixels, not depth or grip points. Lucas: Right. For decades, the industry tried to solve this with rigid programming. You tell the robot exactly where the object is, and it goes there. But in a bin, the object moves. It shifts. It gets buried. To fix this, companies are now deploying deep learning models trained on millions of images of cluttered bins. They are teaching the robot to recognize not just what an object is, but how it sits in space relative to its neighbors. Luna: Is it more about the software seeing better, or the hardware gripping better? Lucas: It is both, but the software is leading the charge right now. The key breakthrough has been in point cloud processing. Instead of just 2D images, the robot uses LiDAR or stereo cameras to build a 3D map of the bin in real time. It identifies potential grasp points on each object, simulates the move virtually, and only commits if the simulation shows no collision with nearby items. It is essentially playing chess with physical objects before it makes its first move. Luna: That sounds computationally expensive. Does it slow down the picking speed? Lucas: It used to. A few years ago, calculating that path took seconds per item. That is too slow for high-volume e-commerce. But with specialized AI chips designed for inference, they can now process the point cloud and generate a grasp plan in milliseconds. The bottleneck has shifted from computation to mechanical execution. The robot can think fast enough; now it needs to move precisely enough. Luna: So we are moving from brute force strength to surgical precision. Lucas: Precisely. And this brings us to the gripper technology itself. Standard parallel jaw grippers are useless here. If you try to pinch a soft bag of chips or a tangled cable, it slips or crushes. Companies are investing heavily in adaptive grippers. These use soft materials, like silicone fingers with internal air chambers, that conform to the shape of the object. They provide a larger surface area for friction, making it easier to lift irregular items without crushing them. Luna: Like a human hand wrapping around a mug instead of clamping it with pliers. Lucas: That is a good analogy. But it is not just about softness. It is about sensor integration. Modern robotic hands are covered in tactile sensors. They can feel slip. If the object starts to slide during the lift, the sensors detect the micro-movement and adjust the grip pressure instantly. This closed-loop control is critical. Without it, even a perfect visual plan fails the moment physics intervenes. Luna: It feels like we are finally getting past the hype cycle into actual utility. Lucas: We are. And the economic case is becoming undeniable. Labor costs in warehousing are rising by four to five percent annually. Automation capital costs are falling. When you combine reliable bin picking with fast cycle times, the return on investment flips within eighteen to twenty-four months. That is short enough for CFOs to approve these projects aggressively. Luna: So which companies are actually doing this well right now? Lucas: There are a few clear leaders. UiPath and other software integrators are partnering with hardware firms like Fetch Robotics and Locus Robotics. But the pure-play robotics companies are also stepping up. KUKA and FANUC, traditionally industrial stalwarts, are releasing new collaborative arms specifically designed for loose-part handling. They are lighter, safer to work near humans, and equipped with the advanced vision systems we just discussed. Luna: Collaborative arms make sense because they don't need safety cages, so they can work right next to the packing stations. Lucas: Yes. In a traditional factory, robots were isolated behind fences. In a flexible warehouse, they need to be everywhere. Cobots allow you to deploy a robot at every packing bench. One robot picks the item from the bin, another places it in the box, and a third seals it. The entire flow becomes continuous and automated, reducing the need for human intervention to nearly zero. Luna: Does this mean fewer jobs in warehouses, or just different jobs? Lucas: It means different jobs. The low-skill repetitive lifting disappears. But you need technicians to maintain these complex systems, data analysts to optimize the picking algorithms, and supervisors to manage the hybrid human-robot workflow. The skill set required shifts from manual labor to technical oversight. It is not a job killer; it is a job transformer. Luna: I suppose that is better than the alternative of constant turnover due to burnout. Lucas: Absolutely. Human workers get tired. Their reaction times slow down. Their attention drifts. A robot does not. However, there is a limit to what a robot can do. It struggles with items that have no defined structure, like a pile of clothing or a bag of mixed small parts. For those, humans are still superior. So the ideal model is a hybrid system. Luna: Where the robot handles the heavy, repetitive lifting, and the human handles the nuanced exceptions. Lucas: Exactly. The robot is the muscle; the human is the brain for edge cases. This division of labor maximizes throughput while minimizing error rates. It is a pragmatic approach rather than a purely technological one. And it is working. Luna: If today's tech conversation gave you something usable, the way these stay ad-free is listener support — buy me a coffee dot com slash fexingo. Lucas: Yeah, if you found value in understanding how the mechanics of automation are shifting, consider supporting the show. It helps us keep digging into these details without relying on corporate sponsors who might want us to soften our critiques. Luna: It keeps the coverage honest and the production quality high. Thanks for listening. Lucas: Back to the technical side. One major hurdle remaining is lighting. Warehouses are often dimly lit or have harsh shadows that confuse optical sensors. How are engineers dealing with that? Luna: Do they use infrared or structured light to bypass visible light issues? Lucas: Both. Structured light projects a known pattern onto the scene, allowing the camera to calculate depth based on distortion, regardless of ambient light. Infrared is also becoming standard because it penetrates dust and smoke better than visible light. This ensures consistent performance across different shift changes and warehouse zones. Luna: Consistency is key. You cannot have a robot failing only at night. Lucas: Right. Reliability is the product. If a robot fails ten percent of the time, it creates more work than it solves. The target failure rate for these systems is below one percent. Achieving that requires redundant sensing. Visual data plus tactile feedback plus inertial measurement units. If one sensor disagrees with the others, the system hesitates or asks for human help. Luna: So it is building in fail-safes rather than trying to be perfectly autonomous all the time. Lucas: Correct. Perfect autonomy is a myth in complex environments. Graceful degradation is the reality. The system works best until it hits a weird edge case, then it politely steps back and lets a human take over. That humility is what makes it viable. Luna: It sounds like we are moving toward a world where machines handle the mundane, leaving humans to handle the unexpected. Lucas: That is the goal. And as the AI models get better at predicting these edge cases, the human role will shrink further. But for now, the partnership is essential. Luna: Do you think smaller warehouses can afford this transition, or is it just for the giants? Lucas: Cloud-based robotics platforms are changing that. Instead of buying expensive proprietary software, smaller facilities can rent computing power and AI models on a subscription basis. This lowers the barrier to entry significantly. You do not need a team of PhDs to run your warehouse anymore. You just need a good internet connection and the right hardware. Luna: Democratizing the technology. That makes sense. Lucas: It does. And it accelerates adoption. As more warehouses adopt these systems, the collective data pool grows, making the AI smarter for everyone. It is a network effect in reverse. More users make the tool better for all users. Luna: So the bin picking problem is not just solved; it is being solved faster by sharing data. Lucas: Exactly. Collaboration between competitors in data sharing is rare, but in logistics, survival depends on efficiency. They are forced to cooperate on the technical front to compete on the service front. Luna: A fascinating dynamic. Efficiency drives cooperation. Lucas: It is. And as we look ahead, the next frontier is not just picking, but sorting and packing simultaneously. Single-arm systems that can inspect, pick, and place in one fluid motion. It reduces cycle time further and simplifies the mechanical design. Luna: One arm to rule them all. Sounds ambitious. Lucas: Ambitious, but likely. The technology is converging. Vision, AI, and mechanics are all improving in tandem. We are close to a tipping point where bin picking is cheaper and faster than human sorting for most standard SKUs. Luna: When do you think that tipping point arrives? Lucas: For large volumes, it is already here. For small batches, maybe three to five years. The question is not if, but how quickly the industry adapts to the new normal. Luna: Adapting to the chaos, one box at a time. Lucas: One box at a time. Until the whole pile disappears. Luna: Will we ever see a fully dark warehouse, where no lights are needed because the robots use their own sensors? Lucas: We might. If they switch entirely to active sensing like radar or sonar, visibility becomes irrelevant. But energy costs for lighting are trivial compared to labor. So unless there is a security reason, full darkness is unlikely. But it is a cool thought experiment. Luna: It is. Like a sci-fi novel coming true. Lucas: Perhaps. But for now, let's stick to the practical improvements. The incremental gains are where the real value lies. Luna: Incremental wins add up. Especially in margins. Lucas: They do. And in logistics, margin is everything. Every cent saved on picking is a cent gained in profit. Luna: So the robots are not just cool gadgets; they are balance sheet tools. Lucas: Exactly. And that financial pressure is the engine driving all this innovation. It is not vanity; it is necessity. Luna: Necessity is the mother of invention, indeed. Lucas: It certainly is. And it is producing some remarkable solutions. Luna: What do you think is the next big leap after bin picking? Lucas: Maybe autonomous mobile robots navigating crowded spaces without collisions. That is the next logical step. Moving the boxes efficiently is just as hard as picking them. Luna: Crowded floors are tricky. Humans are unpredictable. Lucas: Humans are very unpredictable. But the AI is learning to predict human movement patterns. It is getting better at anticipating where a person will step next. Luna: So the robots start to anticipate us, rather than just reacting to us. Lucas: Yes. That is the shift from reactive to proactive automation. It feels much smoother and safer. Luna: I like the sound of that. Safer, smoother, and more efficient. Lucas: It is a significant evolution. And we will see more of it in the coming quarters. Luna: Until then, we watch the boxes get picked. Lucas: We do. And we wonder what the next pile will look like.