I love how bad robots are at picking stuff up. It's amazing how far science has advanced but the act of picking something up that hasn't been seen before is almost impossible to do reliably. It really helps illustrate how AI is really not as powerful as people think it is.
Yeah. You (I assume, unless you are a dog on the internet) and I (definitely not a dog) are the result of something on the order of billions of iterations of the "grasping hand" problem.
It's not even just picking something up, it's picking one specific item out of an arbitrarily messy or overstuffed bin without dislodging or dropping anything else. Items in a bin can be wedged anywhere, in any position, in nondescript boxes only identifiable by the ASIN, or a vague natural language description. It's a hard visual logic, inference and manual dexterity problem made harder by unpredictability and harder still by the need to meet quotas, and every job has its own mess of them.
And on top of that, employees in Amazon warehouses are often cross-trained to save costs, which means you either need one robot that can unload, decant, move juice carts, pick, stow, pallete-stow, etc, at a moment's notice or else multiple dedicated robots for each task.
It's something humans do so easily that Amazon can justify paying almost nothing for the work, but still way beyond what AI and robotics can achieve. The Kiva pods seem to be the current state of the art for Amazon warehouses, and all they seem to do is move bins around (and can't even get that right sometimes.)
This is an important point. Grasping itself isn't that difficult if you know the object and its orientation. It's a pre-configured action (grip here and here, move, release). Ocado, a large online-only grocery store in the UK have a demo they show trade shows that uses a suction cup. They don't bother with a grasper, they just locate a big enough surface to latch onto and use that. That's considerably easier than picking things up using fingers and you can use fairly simple algorithms to detect planar surfaces.
The problems start when things are tightly packed. One object might be in the way of your gripper so you can't take the thing you want. What happens if you drop the object? Does your robot know how to move things aside to grab the item you want?
And on top of this you have the whole issue of detecting objects in a cluttered environment: this requires state of the art semantic segmentation, robust 3D vision that will work on non-cooperative targets and a classifier that can work with occlusions and huge numbers of objects reliably.
I think it's reasonable that some items can be scanned upon entry into the warehouse, e.g. on a laser triangulation turntable that also grabs a video as it rotates. With this sort of task you want to exploit as many priors as you can find. If you look at papers (and factories), you see a lot of tricks: objects on flat surfaces that make segmentation easier, small numbers of known objects, etc.
>I think it's reasonable that some items can be scanned upon entry into the warehouse, e.g. on a laser triangulation turntable that also grabs a video as it rotates
If it could be done quickly enough with enough volume, maybe, but the constraint is that items have to be loaded off the truck, scanned and stowed into a bin (which makes them available for purchase on the site) as quickly as possible.
You could have a conveyor system which would cover 5/6 directions without too much trouble, and it'd be fast enough for real-time capture. Or you could automate the most common purchases e.g. Amazon basics items, Kindles, electronics. They could even ask sellers to provide models or dimensions.
Probably not an issue for them, Ocado sell pretty much only food so it'd be for boxes/bags mostly. It's not been deployed yet though. I think they're also testing gripper robots for picking up larger items like bottle packs.
Also books, unless they're shrink wrapped, have a tendency to open which I imagine would be a pain.
It's incredible how versatile our hands are. Simple pressure, two jawed chuck, three jawed chuck, four jawed chuck, roundgrip, tweezer grip, scoop and all that in a tiny, totally quiet package with force feedback across the whole surface.
Mimicking that mechanically is a major challenge, it has nothing to do with AI per se.
It starts to click when you think about how many years it takes us to form those basic skills, refine them, and then the years spent practicing specific applications. All of this still requires dedicated trainers in the form of parents, driven by love and commitment.
Not even just on an individual basis. If you think about how long it took for dexterous hands to develop on an evolutionary basis, it took tens/hundreds of millions of years or even billions of years depending on where you start from.
Also, there is an interesting theory that humans hands not only evolved to grab things, human hands evolved to form "efficient" fists for punching.
For what it's worth I for example only learned to tie my shoes when I was already 9 years old, while I had learned to read almost all by myself at 5 (my dad had taught me the letters of the alphabet).
Even if I restrict myself to two fingers clamping around something, I can do ridiculously better than the robot. You could even give me the actual robot's hand on a grip. The control logic seems to be the much bigger problem to me.
Force feedback provides humans constant information on how our hands are doing.
Reliable force feedback data at our hand's level of detail would provide a wealth of data to those making the control software and that better software could benefit from real time data. The two problems are strongly interconnected.
When we are small we barely do any better than current gen robots.
Thing is that each of us accumulate prior data that we can use as a basis for future action, things like how much pressure is enough for grasping a raw egg without breaking it.
I totally agree. (I only disagreed with the tiniest detail of his point and meant only to rebut that).
I would like to point that with pair of tongs you will still have force feedback. Even though is will be reduced in precision it will still be better than any artificial system I am aware.
Not only that, but it's difficult to mathematically model deformable objects in ways that are computationally tractable (e.g. running on a 10ms update loop)
If you think you've got a really good solution for robotics grasping you shouldn't be competing in the Amazon Picking Challenge given the IP terms. We (Righthand Robotics) think we've got a handle on the problem but didn't want to get involved for that reason. There are some other companies out there with interesting solutions who weren't involved either.
People are competing in the Amazon Picking Challenge too! But the picking prize is just $50,000. If it were $1,000,000 like the Netflix prize I'm sure there would be more competition.
Actually there are a couple of companies and labs who are close to (if not already solved) the problem, but no sane person is going to compete in the Amazon Picking Competition if they've actually solved it.
I find it interesting that they try so hard to eliminate warehouse workers, all the while making structural and capital decisions that ultimately increase inventory, shipping, planning, procurement, split shipment, and reverse logistics costs...all of which add up to way more of a cost problem than the handful of cents per unit of warehouse processing costs. Amazon's cost problems have almost nothing to do with warehouse workers and everything to do with management and their inability to comprehend the consequences of their rash decisions. And because their capital plan is consumed by these bad decisions, they are quite literally locking them into place.
I'm increasingly convinced that amazon has opened themselves up to competition on simultaneous price and performance fronts, even with relatively few robots and a unionized and well paid work force. A few key structural decisions made up front by intelligent and informed people will make all the difference.
"capital decisions that ultimately increase inventory, shipping, planning, procurement, split shipment, and reverse logistics costs."
That's interesting! What are these decisions? I'm guessing there is a natural increase due to Amazon's growth but are they making any structural changes to make the system more unwieldy?
1) Inventory growth. Inventory turns have decreased every year (look at their public statements!), to the point where they are worse than brick and mortar retail. Some inventory growth is good...you need to buy more stuff to sell more stuff. But a healthy company would see increasing inventory turns over time, leveling off once they figure out what items are worth stocking and what are not.
2) Warehouse growth. The underlying assumption is that if you're closer to the customer, you save more money on transportation. That's true to some extent, but they fuck it up in other ways: fragmenting inventory resulting in split shipments (a way bigger deal than most people realize, I wouldn't be surprised if this alone overwhelmed the savings from shorter linehauls), increasing safety stocks, truck endpoint combinatorial explosions, truck scheduling issues, inventory balancing problems, reduced truckload utilization factors from reduced shipping volumes per warehouse and destiation. Larger warehouses have their problems, but compared to the issues that more warehouses cause, they are very solvable in comparison.
3) Random stow processes and path-based picking are hard wired into everything about their operations. From their software to their warehouse design to their metrics. It works phenomenally for books, but not so well for things that aren't physically shaped like books or have sales distributions like books. Inbound costs with random stow are orders of magnitude higher than case-level stow for case-packed items, experience exponential growth in costs when inventory shelves are at high levels of utilization, and random stow combined with path-based picking means your pick rates decrease almost linearly with inventory turns (more inventory per unit sold means walking farther between each pick which means more labor cost per pick, and prioritized placement to fix the pick cost problem just shifts costs towards a much more complex and costly stow process).
4) Refusal to reevaluate decisions made in the past that no longer apply. Two big examples are random stow processes, and inventory selection.
- Random stow vs other stow methods were evaluated back when Amazon was a bookstore, and they were an astounding win. Now that Amazon sells everything, there are plenty of methods that are vastly superior, and even the possibility of different methods for different products, but they've built out several hundred warehouses under the assumption of random stow and would require monumental efforts to retool them and write the software to support it.
- Inventory selection is seen as a universal good and is never going to stop growing. That's fine for a business growing into new sectors and categories, but when you have 3000 different brands and models of staplers, you've passed the point of marginal benefits exceeding marginal costs. The effect on inventory turns, warehouse growth, and warehouse operations alone is staggering.
5) Their Kiva robots work beautifully in small warehouses with low shipping volumes, but have severe scaling issues in larger warehouses due to inherent design problems. They have centralized path planning which is an NP-complete problem that can't keep up with the real time requirement at large warehouses. They require whole-warehouse shutdowns every time a single piece of inventory falls off a shelf. They experience physical deadlocks due to inventory being tied up in multiple conflicting processes at the same time. Sure, they eliminate picking labor, but they replace it with decreases in stow and packing productivity, and unseen to it all is that the fact that dependence on Kiva means small warehouses which means you further explode the logistics problem by having more warehouses and more fragmented inventory.
6) They introduced and scaled FBA without item-level inventory accounting. Now, due to co-mingling, fraud and piracy are a systemic problem. And it's not entirely solvable without item level accounting...which would require phy...
> They require whole-warehouse shutdowns every time a single piece of inventory falls off a shelf.
I'm currently working in an 800,000 foot Amazon warehouse that uses Kivas, where inventory falls out of the bins constantly, and they don't even shut the whole thing down if the pods crash into each other. They just shut off a section of the pods and send someone in to clean the mess up.
I'm not saying there aren't problems - I could go on a rant about the issues I've seen with their internal software and UI design, not to mention bin design and the way it seems pods are allocated - but it doesn't seem to be that broken.
20 cents per item just to pick it out of a bin and pack it in a box has to be a significant chunk out of the profit Amazon can make from those smaller low value items that people order individually with Prime.
Amazon does not care if one or two orders represent a loss. As long as you move more and more of your purchases, hell people shop their _groceries_ online, to Amazon, they stand to profit from locking you in.
Amazon wants your first thought on the question "I want XYZ, how do I get it" to be "Amazon".
It's amazing how bad robot manipulation is in unstructured environments. It hasn't improved a lot in 40 years. There's was a DARPA manipulation challenge in 2012.[1] Simple manipulation is slow and clunky, even with really good robot hands.
It's interesting because the AI founders like McCarthy considered the intellectual challenges to be hard, and the robot challenges to be relatively easy.
They would have probably been really surprised that we'd have a superhuman Go player before we had a robot that was just average at picking up arbitrary items.
A good part of that is that the kid's nervous system is continually being detuned due to growth. Consider that dog learns how to catch a ball reliably once it has reached full size.
Interestingly the myelin sheath around your peripheral nerves thins as you grow. I have always assumed this is to reduce capacitance so that signal propagation time between (say) your fingertips and brain remains roughly constant as your arm gets longer. I don't know if anyone has ever studied this though.
You may be right that thinning during maturation could serve a purpose, but if so you have the effect backwards: thinner myelin should mean higher capacitance (C ~ A/d) and slower conduction velocity. One would expect, if anything, that the myelin would need to get thicker as the limb lengthens. Perhaps nerve fiber diameter also increases with maturation---which also results in faster conduction velocity---and the myelin sheath thins to compensate.
Part of it is that training the superhuman Go player is more parallelizable. The challenge is entirely virtual, and you can have unlimited synthetic training data by having Go-playing algorithms play each other.
I've never worked with machine learning in robotics, but my understanding is that physics simulators are still insufficient environments for training state-of-the-art robotic agents. Especially when using reinforcement learning or a similar algorithm, many models are trained with several real robots picking up real objects.
Ironically Marvin Minsky discouraged me from working on robotics because it was so easy to get distracted by the hard problems in robotics and stop working on the hard problems in AI. He supported this argument through all the experience he had had building robot arms and the like.
It really seemed unfair that he had had the fun of building the robots and than said I shouldn't do it. It took years to realize what he was really saying (a statement I could make about many things he said).
Grasping Robots Compete to [be minimally competent at picking stuff up and after many years trying are still hilariously bad compared to the humans that] Rule Amazon's Warehouses.
They are getting better these days, but let's not get too breathless about it. That headline is very misleading.
57 comments
[ 3.1 ms ] story [ 121 ms ] threadIt's not even just picking something up, it's picking one specific item out of an arbitrarily messy or overstuffed bin without dislodging or dropping anything else. Items in a bin can be wedged anywhere, in any position, in nondescript boxes only identifiable by the ASIN, or a vague natural language description. It's a hard visual logic, inference and manual dexterity problem made harder by unpredictability and harder still by the need to meet quotas, and every job has its own mess of them.
And on top of that, employees in Amazon warehouses are often cross-trained to save costs, which means you either need one robot that can unload, decant, move juice carts, pick, stow, pallete-stow, etc, at a moment's notice or else multiple dedicated robots for each task.
It's something humans do so easily that Amazon can justify paying almost nothing for the work, but still way beyond what AI and robotics can achieve. The Kiva pods seem to be the current state of the art for Amazon warehouses, and all they seem to do is move bins around (and can't even get that right sometimes.)
The problems start when things are tightly packed. One object might be in the way of your gripper so you can't take the thing you want. What happens if you drop the object? Does your robot know how to move things aside to grab the item you want?
And on top of this you have the whole issue of detecting objects in a cluttered environment: this requires state of the art semantic segmentation, robust 3D vision that will work on non-cooperative targets and a classifier that can work with occlusions and huge numbers of objects reliably.
I think it's reasonable that some items can be scanned upon entry into the warehouse, e.g. on a laser triangulation turntable that also grabs a video as it rotates. With this sort of task you want to exploit as many priors as you can find. If you look at papers (and factories), you see a lot of tricks: objects on flat surfaces that make segmentation easier, small numbers of known objects, etc.
If it could be done quickly enough with enough volume, maybe, but the constraint is that items have to be loaded off the truck, scanned and stowed into a bin (which makes them available for purchase on the site) as quickly as possible.
Simple algorithm for that : Annoy (pun intended) towers :-)
I wonder how they pick up books using this method?
Books are difficult because they appear to have hard surfaces on the side, which turn out to be not so hard and not suitable for gripping.
Also books, unless they're shrink wrapped, have a tendency to open which I imagine would be a pain.
Mimicking that mechanically is a major challenge, it has nothing to do with AI per se.
It's not a trivial challenge in any form really.
Also, there is an interesting theory that humans hands not only evolved to grab things, human hands evolved to form "efficient" fists for punching.
http://www.latimes.com/science/sciencenow/la-sci-sn-human-fi...
For what it's worth I for example only learned to tie my shoes when I was already 9 years old, while I had learned to read almost all by myself at 5 (my dad had taught me the letters of the alphabet).
Reliable force feedback data at our hand's level of detail would provide a wealth of data to those making the control software and that better software could benefit from real time data. The two problems are strongly interconnected.
Thing is that each of us accumulate prior data that we can use as a basis for future action, things like how much pressure is enough for grasping a raw egg without breaking it.
I would like to point that with pair of tongs you will still have force feedback. Even though is will be reduced in precision it will still be better than any artificial system I am aware.
http://www.takktile.com/main:tech
There are other methods of figuring out how much force a hand is exerting too.
EDIT: Pun retroactively intended.
This makes me wonder. Why did people compete for the Netflix prize, [1]? Were those IP terms better?
[1] https://en.wikipedia.org/wiki/Netflix_Prize
I'm increasingly convinced that amazon has opened themselves up to competition on simultaneous price and performance fronts, even with relatively few robots and a unionized and well paid work force. A few key structural decisions made up front by intelligent and informed people will make all the difference.
That's interesting! What are these decisions? I'm guessing there is a natural increase due to Amazon's growth but are they making any structural changes to make the system more unwieldy?
2) Warehouse growth. The underlying assumption is that if you're closer to the customer, you save more money on transportation. That's true to some extent, but they fuck it up in other ways: fragmenting inventory resulting in split shipments (a way bigger deal than most people realize, I wouldn't be surprised if this alone overwhelmed the savings from shorter linehauls), increasing safety stocks, truck endpoint combinatorial explosions, truck scheduling issues, inventory balancing problems, reduced truckload utilization factors from reduced shipping volumes per warehouse and destiation. Larger warehouses have their problems, but compared to the issues that more warehouses cause, they are very solvable in comparison.
3) Random stow processes and path-based picking are hard wired into everything about their operations. From their software to their warehouse design to their metrics. It works phenomenally for books, but not so well for things that aren't physically shaped like books or have sales distributions like books. Inbound costs with random stow are orders of magnitude higher than case-level stow for case-packed items, experience exponential growth in costs when inventory shelves are at high levels of utilization, and random stow combined with path-based picking means your pick rates decrease almost linearly with inventory turns (more inventory per unit sold means walking farther between each pick which means more labor cost per pick, and prioritized placement to fix the pick cost problem just shifts costs towards a much more complex and costly stow process).
4) Refusal to reevaluate decisions made in the past that no longer apply. Two big examples are random stow processes, and inventory selection.
- Random stow vs other stow methods were evaluated back when Amazon was a bookstore, and they were an astounding win. Now that Amazon sells everything, there are plenty of methods that are vastly superior, and even the possibility of different methods for different products, but they've built out several hundred warehouses under the assumption of random stow and would require monumental efforts to retool them and write the software to support it.
- Inventory selection is seen as a universal good and is never going to stop growing. That's fine for a business growing into new sectors and categories, but when you have 3000 different brands and models of staplers, you've passed the point of marginal benefits exceeding marginal costs. The effect on inventory turns, warehouse growth, and warehouse operations alone is staggering.
5) Their Kiva robots work beautifully in small warehouses with low shipping volumes, but have severe scaling issues in larger warehouses due to inherent design problems. They have centralized path planning which is an NP-complete problem that can't keep up with the real time requirement at large warehouses. They require whole-warehouse shutdowns every time a single piece of inventory falls off a shelf. They experience physical deadlocks due to inventory being tied up in multiple conflicting processes at the same time. Sure, they eliminate picking labor, but they replace it with decreases in stow and packing productivity, and unseen to it all is that the fact that dependence on Kiva means small warehouses which means you further explode the logistics problem by having more warehouses and more fragmented inventory.
6) They introduced and scaled FBA without item-level inventory accounting. Now, due to co-mingling, fraud and piracy are a systemic problem. And it's not entirely solvable without item level accounting...which would require phy...
I'm currently working in an 800,000 foot Amazon warehouse that uses Kivas, where inventory falls out of the bins constantly, and they don't even shut the whole thing down if the pods crash into each other. They just shut off a section of the pods and send someone in to clean the mess up.
I'm not saying there aren't problems - I could go on a rant about the issues I've seen with their internal software and UI design, not to mention bin design and the way it seems pods are allocated - but it doesn't seem to be that broken.
https://news.ycombinator.com/item?id=14932427
Amazon wants your first thought on the question "I want XYZ, how do I get it" to be "Amazon".
[1] https://www.youtube.com/watch?v=jeABMoYJGEU
They would have probably been really surprised that we'd have a superhuman Go player before we had a robot that was just average at picking up arbitrary items.
Interestingly the myelin sheath around your peripheral nerves thins as you grow. I have always assumed this is to reduce capacitance so that signal propagation time between (say) your fingertips and brain remains roughly constant as your arm gets longer. I don't know if anyone has ever studied this though.
Myelination continues Into adulthood btw -- you can see a change in the distribution of white matter due to learning a new language for example.
I've never worked with machine learning in robotics, but my understanding is that physics simulators are still insufficient environments for training state-of-the-art robotic agents. Especially when using reinforcement learning or a similar algorithm, many models are trained with several real robots picking up real objects.
It really seemed unfair that he had had the fun of building the robots and than said I shouldn't do it. It took years to realize what he was really saying (a statement I could make about many things he said).
They are getting better these days, but let's not get too breathless about it. That headline is very misleading.
https://www.youtube.com/watch?v=l8zKZLqkfII