Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Pack a Suitcase
Transcript
- Lucas: You know that moment when you're packing a suitcase and you realize a pair of shoes and a toiletry bag just don't fit together the way you thought they would? Luna: I spend a solid fifteen minutes re-arranging every single time. It's like a spatial puzzle. Lucas: That puzzle turns out to be one of the hardest unsolved problems in robotics. I'm not exaggerating. Packing a suitcase — loading arbitrarily shaped, deformable, fragile objects into a container — is still something no general-purpose robot arm can do reliably. Luna: We've done episodes on folding towels, picking apples, even peeling potatoes. How is packing a suitcase different? Lucas: Great question. Folding a towel is a known geometry — you just have to track the corners. Picking an apple is a known shape with a known gripper strategy. But packing a suitcase? The robot has to reason about a three-dimensional arrangement of objects it may have never seen before, with varying stiffness — a shirt compresses, a phone charger doesn't — and it has to decide an order that won't cause the stack to topple or crush fragile items. Luna: So it's the combinatorial optimization combined with real-world physics. Lucas: Exactly. Computer scientists call this the three-dimensional bin packing problem, and it's np hard — meaning the time to find an optimal solution grows exponentially with the number of items. But humans solve it intuitively with heuristics: put heavy things at the bottom, fill gaps with soft things, arrange like Tetris pieces. Luna: And robots just don't have that intuition. Lucas: Not yet. In 2024, researchers at UC Berkeley and NVIDIA published a paper where a robot arm tried to pack a random set of 12 objects into a tote. On average, it took over two minutes per object — and that was in simulation with perfect sensor data. In the real world, with a noisy camera and a slightly off-calibration gripper, the success rate dropped below forty percent. Luna: Two minutes per object sounds glacial for a warehouse. Lucas: It is. Amazon's Sparrow robot — the one they use for picking items out of bins — can grab a product in about two seconds. But packing it into a shipping box in a stable, space-efficient way? That's still largely done by humans. Amazon has said publicly that the packing problem is one of their top robotics bottlenecks. Luna: And it's not just e-commerce. Think about logistics for moving pallets, loading trucks, arranging laboratory equipment. Lucas: Right. Every time you see a truck being loaded by hand, part of the reason is that a robot can't yet look at an irregular pile and decide how to make it fit efficiently. Boston Dynamics' Stretch robot is designed for unloading trucks — it slides boxes onto a conveyor. But loading? That requires a different level of spatial reasoning. Luna: What makes the problem so stubborn? Is it sensing, planning, or the hardware itself? Lucas: All three, but the core challenge is planning under uncertainty. The robot has to estimate the shape and stiffness of each object from a single camera view. A shirt looks like a blob. A pair of headphones has wires that dangle. The planner has to simulate thousands of possible arrangements in its head — but each simulation is only as good as the model of how objects deform. If the model assumes a shirt is a rigid block, the packing will fail. Luna: So they need better physics models. Lucas: Yes. Researchers are using neural networks to learn deformation from visual data — training on millions of simulated grasps. Google's DeepMind recently released a model called 'Suitcase' — no, really, that's what they called it — that can pack a suitcase in simulation with near-human efficiency. But transferring that to a real robot arm with friction, sensor noise, and motor inaccuracies is a massive leap. Luna: That simulation to reality gap keeps coming up in every manipulation episode we do. Lucas: It's the theme of the decade. For packing, the gap is especially wide because small errors compound. If you misplace one item by a centimeter, the next item might not fit at all. And unlike a human, the robot doesn't have tactile feedback across its entire hand — it only feels force at the fingertips. Luna: Quick honest thing — a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that's literally what funds making this many of these deep-dive episodes. If today's tech conversation gave you something usable, that's the reason we can keep spending hours on a single problem. Lucas: Yeah, it's a small group that makes it possible. No ads, no sponsors — just people who want better robotics coverage. And I think it shows in the depth we can go. Luna: Exactly. Alright, back to why packing is still manual. There's also a hardware bottleneck, right? Lucas: Huge one. Most robot grippers are two-fingered parallel jaws. They work fine for picking a box off a shelf. But try picking up a T-shirt with two fingers — it slips, or you pinch it in a way that makes it hard to place. Amazon's Sparrow uses a vacuum gripper with multiple suction cups, which helps with irregular shapes. But suction fails on porous fabric or curved surfaces. Luna: So do you need a more human-like hand? Lucas: Maybe, but that adds complexity. The Shadow Dexterous Hand has 24 joints — controlling them for a packing task is computationally expensive. Some labs are using soft robotic grippers that conform to object shape, but they lack the precision to place items tightly. There's a trade-off between compliance and accuracy that nobody has fully solved. Luna: Let's talk about a real-world deployment. Amazon has a system called 'Universal Item Sorter' that uses robots to sort packages into bins. But that's one item at a time into a fixed bin. Packing a mixed order into one box — like a book, a bottle, and a scarf — is a different beast. Lucas: That's exactly the problem. In 2025, Amazon filed a patent for a robotic packing cell that uses a combination of a six-axis arm, a vacuum gripper, and a vibrating table to settle objects. The idea is to place items loosely, then vibrate the box to shift them into position. But that only works for items that aren't fragile. A wine bottle wouldn't survive. Luna: So they might need different strategies per item type. Lucas: That's where the research is heading. A system called 'PackingNet' from the MIT CSAIL lab uses a neural network that first classifies each object into one of five categories — rigid, soft, fragile, cylindrical, or irregular — then applies a packing strategy specific to that category. In simulation, they achieved a volumetric efficiency of eighty-seven percent, which is close to the human average of ninety-one percent. Luna: Eight-seven percent is impressive. What's the flaw? Lucas: It assumes perfect knowledge of object shape and stiffness. In a real warehouse, the system would need to infer those properties on the fly. And it still takes about thirty seconds per item — fine for a slow logistics line, but not for Amazon's pace. Luna: What about the order of packing? Humans tend to put heavy, rigid items first, then fill gaps with soft items. Does the robot learn that? Lucas: Some do. Researchers at the University of Tokyo trained a reinforcement learning agent to pack a suitcase by rewarding it for stability and volume efficiency. The agent learned to put shoes at the bottom and shirts on top without being explicitly told. But training took six million episodes in simulation — that's about two weeks of compute time. And the policy didn't transfer to real objects with different textures or weights. Luna: So the training is too narrow. Lucas: Exactly. The dream is a foundation model for manipulation — a large neural network trained on diverse packing scenarios that can generalize. I think we'll see something like that within five years, but for now, packing a suitcase remains a human specialty. Luna: It's funny — we have robots that can perform surgery and assemble car engines, but packing a weekend bag is the frontier. Lucas: It's a reminder that the problems that look easy to us are often the hardest for machines. A five-year-old can pack a toy box. A robot arm with millions of dollars of engineering still can't. Luna: So what's the next step you're watching? Any promising demos? Lucas: I'm watching a startup called 'Paxos Robotics' — they're building a packing system specifically for fresh grocery delivery, where items have irregular shapes and some are fragile. They use a soft gripper combined with a 3D scanner and a planner that runs in real-time on an edge GPU. Their demo at the 2026 IEEE Conference showed packing twelve items in about forty seconds with no damage. That's approaching practical speed. Luna: Forty seconds for twelve items is about three seconds per item — that's getting close to human pace. Lucas: Right. And they did it with a mix of apples, boxes of eggs, bottles, and bags of lettuce. If they can get that to work in a real distribution center, it could change how we think about automated fulfillment. Luna: I'll be watching that. For now, I guess I'll keep doing my own suitcase Tetris. Lucas: And you'll probably do it faster than any robot. At least for the next few years.