Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Sort a Mixed Bin of Trash
Transcript
- Lucas: Luna, I want to start with something you and I do every single day without thinking: we take a mixed bin of trash — maybe a plastic bottle, a coffee cup, a banana peel, a dirty napkin — and we toss each item into the right bag or bin. It takes us maybe two seconds per item. Now imagine asking a robot arm to do the same thing, from a bin where everything is piled on top of everything else, covered in residue, crumpled, wet. Luna: And the robot fails. Completely. Lucas: Completely. This is episode 87, and we are still talking about things robot arms cannot do. And today's angle — sorting a mixed bin of household trash — might be the hardest one yet. Because it's not just about picking up one object, which we've covered for months. It's about picking up any object, from a pile of unknown objects, without knowing what's coming next. Luna: And the pile itself is constantly changing. The second you grab something, everything shifts. The objects underneath move, occlude each other differently. So the robot can't even rely on a static scene. Lucas: Exactly. That's the key distinction from the previous episodes. In all those cases — the wet bar of soap, the raw egg, the single sheet of paper — the robot knew what it was trying to pick. The object was isolated, known, often in a fixed position. Here, the robot walks up to a bin and sees a chaotic jumble of maybe fifty items, many of which it has never seen before, made of different materials, in different states of deformation. Luna: And the sensor problem is brutal. Computer vision can identify objects when they're clean and separated, but in a trash bin, everything is overlapping. A bottle might be half-occluded by a napkin. A banana peel might look like a crumpled piece of paper from one angle. Lucas: There's a company called AMP Robotics that has been trying to solve this for years. They build robots for recycling facilities. Their systems can pick out plastic bottles and aluminum cans from a conveyor belt with decent accuracy, but when it comes to an actual mixed bin — the unsorted stream as it comes off the truck — the success rate drops dramatically. I've seen numbers around 60 to 70 percent for the easiest items, and below 30 percent for things like soiled cardboard or thin film plastics. Luna: And 70 percent sounds okay until you realize that a human sorter does 95 percent plus. And the robot slows down as it gets confused. So the throughput per hour is still way below a human line. Lucas: Right. And the problem isn't just vision. It's the gripper. Most robot arms use either a parallel jaw gripper or a vacuum suction cup. But a suction cup can't grab a crumpled napkin, and a parallel jaw can't grasp a slippery banana peel without crushing it. So you need a gripper that can switch strategies in milliseconds based on what it's about to touch. And that's an actuation and sensing loop that is still very much a research problem. Luna: So we need a gripper that can feel. Tactile sensing becomes critical. But even the best tactile sensors — like the ones from SynTouch or the GelSight sensors — are still lab-grade. They're expensive, fragile, and don't stand up to the grease and grime of a real recycling center. Lucas: And that's before we talk about the 'sim to real' problem. Companies like Covariant and Osaro train their picking policies in simulation because it's faster and safer. But a simulated trash bin has perfectly rigid objects, clean surfaces, and neat physics. A real trash bin has sticky residues, deformable objects, and objects that are tangled together. The transfer is terrible. Luna: I think this is the episode where we have to admit that maybe the reason robot arms can't sort trash is that they fundamentally lack the kind of 'common sense' physics that humans have. We don't just see a bottle — we know that if we pinch it near the top, it won't collapse. We know that a napkin will tear if we pull too hard. That knowledge comes from years of manipulating objects. Lucas: And we can adapt instantly. I can pick up a wet paper towel with two fingers if I adjust my grip force, but a robot has to explicitly compute that adjustment. And in a mixed bin, it doesn't have time. So the industry is starting to take a different approach: instead of trying to pick any object, they are redesigning the bin itself. Some companies are pre-sorting the waste into categories — heavy, light, wet, dry — using air classifiers and trommel screens, so the robot only has to handle a narrower set of items. Luna: That's a workaround. It's honestly the kind of thing that gives robotics a bad name. 'We can't solve the general problem, so we'll change the environment.' But it's also pragmatic. In a real recycling plant, throughput is king. And if changing the bin design gets you from 30 percent to 80 percent, that's a win. Lucas: It is. And look, if you found this conversation useful — maybe you work in waste management or robotics and this gave you a clearer picture of the bottleneck — that's exactly the kind of thing we try to do here. If today was worth a coffee to you, there's a link: buy me a coffee dot com slash fexingo. No pressure, just a way to keep the show ad-free and focused on the actual engineering challenges. Luna: Yeah, it's a small thing that makes a big difference for us. Now, back to that gripper problem — I want to talk about a specific research paper from MIT's CSAIL last year. Lucas: Which one? There's been a few on dense clutter. Luna: The one where they used a 'fingers' gripper with twelve independent joints, each with a tactile sensor. They trained it in simulation on random piles of geometric shapes, but then they tested it on a real bin of actual trash — coffee cups, crumpled paper, plastic forks. And the success rate for the first grasp attempt was 78 percent. Lucas: Seventy-eight is actually impressive for a research prototype. But I'd bet that number drops after the first item is removed, because the pile settles and the remaining objects are more occluded. And the gripper itself is incredibly complex — twelve joints, each with a motor and sensor — which means it's expensive and hard to clean. Luna: Exactly. So the trade-off is between dexterity and robustness. A simpler gripper like a suction cup is cheap and can be replaced easily, but it fails on porous or irregular objects. A complex gripper handles more but costs more and breaks more often. In a real plant running 24/7, downtime is the enemy. Lucas: That's why we haven't seen widespread deployment of trash-sorting robots, despite hundreds of millions in venture funding. The economic case is still marginal. Labor for sorters is relatively cheap in many regions, and the robots require constant supervision and maintenance. The break-even point, by most estimates, is around three to four years. And that's assuming the robot runs at 80 percent of human speed, which it doesn't. Luna: But there is one place where the economics are shifting: countries with high labor costs and strong recycling mandates, like Japan and South Korea. They are investing heavily in automated sorting because they simply can't find enough human workers for the job. And they are willing to accept lower performance if it means consistent operation. Lucas: So the future might be a hybrid facility: humans handle the difficult edge cases — like sorting out hazardous waste or untangling wires — while robots handle the high-volume, repetitive picks. That's actually happening at some facilities run by ZenRobotics in Europe. They have a robot that picks construction waste — wood, metal, concrete — from a conveyor, and it works because the objects are larger and more distinct. Luna: Household trash is a different beast. It's small, varied, and dirty. And that's why, as of July 2026, the mixed bin remains essentially unsolved. The closest we have is the AMP Cortex system, which uses a combination of computer vision and a suction gripper, and it can hit about 50 picks per minute on a conveyor belt with pre-sorted material. But from a bin? Maybe 10 picks per minute, with a 20 percent failure rate. Lucas: And that failure matters. If the robot mis-sorts a plastic bottle into the paper stream, that batch of paper gets contaminated and may be rejected by the mill. So the cost of error is high. Human sorters are still the gold standard because they can correct themselves in real time. Luna: So what would it take to solve the problem? A breakthrough in tactile sensing that's cheap and durable? A new gripper design? Better sim to real transfer? Or maybe a new approach to object recognition that doesn't require a perfect model? Lucas: I think it's all of the above, but the biggest gap might be in 'affordance' reasoning — understanding what actions an object affords. A bottle affords being grasped by the neck, a banana peel affords being pinched at one end, a napkin affords being flattened and then lifted. That kind of reasoning is still very much in the research phase. But there are groups at Stanford and Berkeley working on it, using large language models to give robots semantic cues. 'This is a banana peel, so try picking it from the stem end.' Luna: But that requires the robot to first identify the object, which brings us back to the occlusion problem. It's a chicken and egg loop. You need to see to know what to grasp, but you need to grasp to see what's underneath. Lucas: And that's the core challenge. Until we solve that, robot arms will continue to fail at sorting a mixed bin of trash. It's a humbling reminder of how far we are from human-level dexterity. Luna: And a reminder that even the most advanced industrial robot is still a long way from replacing a minimum-wage worker with a pair of hands and a few years of experience. Lucas: So that's where we stand as of mid-2026. The mixed bin problem is still open. If you're working on it, we'd love to hear from you. And if you just learned something, maybe that's worth a coffee. Same link: buy me a coffee dot com slash fexingo. Thanks for listening.