Interesting to compare "build an environment and process so we can use simple robots in it" to the Boston Dynamics - "build robots that can work in complex environments".
Also interesting to see the robot arm being used to move the products that currently come in boxes (suitable for humans to unpack) if robots become the norm will we see them being distributed like electronic components on reels/tape that are easier to unpack mechanically
In one of Asimov's books, The Caves of Steel, he talked about how Androids run all of the farming equipment on Earth to feed everyone. I don't remember the exact line, but the reasoning for Androids running regular farm equipment was that, in his universe, it was just way easier and cheaper to build Androids that can use all the same tools humans can, instead of building automated versions of all the tools.
The other interesting tidbit was that the vast majority of the world was farm land. Humans lived in huge cities with tons of verticality. This was because there were too many humans on Earth - a whopping 9 Billion.
How could he be so imaginative and even prescient for some things, and yet think that at 9B the world would be overpopulated? I mean, if you’re playing with that concept why not crank it up to 20 or 30?
Caves of Steel was published in 1953. Back in those times, population dynamics weren't as widely known. The world had just passed 2.5 billion and naive rates would put it at 9 billion over a hundred years later.
It seems like old science fiction authors regularly set the time of their worlds much closer to now than reasonable for some reason. To be more relatable perhaps? Blade Runner was set in 2019 for example: fully synthetic humans, flying cars, wars in space. None of that happened. Then there's 1984 by Orwell, and I'm sure dozens of other examples if you looked. Still waiting for that hoverboard (Back to the Future 2 was set in 2015).
Yes, modules are typically available either on tape&reel or on trays (also easy to pick).
Also the Raspberry Pi pico is an example of a PCB that can be either a final dev board or a module and comes on a Reel: https://makerbright.com/raspberry-pi-pico-reel-bulk.html
When I got my first Raspberry Pi Picos, I got three in a row cut out from a reel. I don't think you'd get ATX motherboards this way, but I do see smaller boards come that way nowadays.
> AutoStore says its first supplied technology to Ocado as early as 2012 and this is the foundation on which the Ocado Smart Platform was built. The company has claimed several patent infringements relating to the design and lifting mechanism of Ocado’s robots.
If true, yikes, but at the same time I have doubts if some aspects of the technology are even patentable (or at least make the claims too narrow to the point that patent-free variations allow workarounds and make the patents useless).
Given that Autostore's patent on the idea is from 1997 [1] and has expired, it's difficult to see how they could have another, non-expired patent on the same thing.
Unless its a copy of the technology, is it protected ip? Its an idea (use tracked robots on top of product bins) and there's some issue about patenting ideas vs technology.
They have automatic bin picking! Amazon hasn't been able to get that to work in production. Amazon's system has Kiva robots moving bins around, but humans pick from the bins.
Expect a big drop in fulfillment center workers when Amazon gets bin picking that works.
Walmart compelled vendors to put barcodes on product in the late 70s early 80s. I'm surprised Amazon hasn't encouraged vendors to supply "robot friendly" or "frustration free" packaging.
Having designed one I can tell you it's not that hard.
Thinking from first principles helps. First, simplify the problem domain. The challenges are generally oddball shapes (think: heavy golf club), heavy items, speed, spatial efficiency, specialist handling requirements, environment, reliability, mobility and cost. If you have unlimited patience (retries allowed, no hard speed requirement) and less limited space (more power, better viewing angles, opportunity for stereoscopic vision), roughly similar form factor SKUs (=CPG), a purpose-built environment and no special handling requirements (nuclear waste/fresh food/fast-melting ice cream), no mobility requirement or cost sensitivity, then it's much easier.
Second, simplify handling by batching items in standard containers. What these system vendors never show you is how much they overcharge you for plastic boxes and how much energy is wasted loading, unloading and cleaning them.
Third, a grid-based system is clearly the most spatially efficient, so design around that.
In these systems the true spatial efficiency for most items is probably poor as Dematic Autostore for instance only offer 3 bin sizes. The true spatio-temporal efficiency is probably middling except when considered in parallel for huge order throughput requirements and a large number of SKUs because an untrained primate can pick faster than a bot for arbitrary items in most cases. Capex is very high, so these systems only make sense if you have a guaranteed large-SKU picking problem that won't go away, and you don't care about owning the knowledge to implement it cost-efficiently (eg. you are a manager looking out for #1 instead of the company's long term bottom line).
These things are terrible for the world, however, as they basically enable the accelerated consumption of single use plastics which is the CPG industry as a whole. I predict that in progressive countries such systems will be taxed or outlawed within our lifetimes, probably first in Europe, probably first in Holland as they are so good with agriculture and logistics.
My conclusion is there is a massively over-represented, inexperienced ROS-oriented robotics community in the west and on HN that believes ML + ROS + third party hardware = whole robotics field. They have less experience implementing custom hardware/mechanical thus fail to consider its possibilities.
Yea that did make me smile also, glad some use still being extracted from the Kinect, it was (and still is) an amazing bit of kit. I suspect they have support from MS for their use case rather thank just hacking it.
I remember this one setup in an aquarium where you could move sand around and the Kinect would read how you have the sand formed and apply geological effects to it.
Pile it high, mountain, carve it out, water and eventually sea. It was pretty amazing as it was all projected into the sand in real time and looked just amazing.
Think it was running on some 1080 or 1080-Ti when that was king of graphics cards.
A lot of local science centers / children's museums have this kind of setup somewhere, and more often than not, I've seen a Kinect hanging out somewhere doing the topographic mapping.
It's a versatile piece of hardware! As long as you're not trying to get granular object detection less than a foot away it's able to see features with a fair amount of detail.
This isn't an area that I study, directly, but I've had some experience with other robotic automation systems on a small/medium scale.
Maybe this is common and my experience is at issue, but I found the grid design with the "workers" on top to be really interesting.
My most direct experience was with a tape robot. This wasn't a run-of-the-mill backup library for a small DC, but a huge room on half of a floor of the datacenter that had a 6 axis[0] "bigger than an average adult man" one-arm robot that is usually seen in an automotive factory. It moved on a track the length of the room, grabbed tapes from a library and inserted it into drives; all of which were attached to a mainframe.
This was done with so few sensors, that the tape robot couldn't tell if it successfully grabbed a tape or missed it; indication of failure wasn't realized until the drive reported no tape present when the operation completed.
When it missed, the tape often ended up in the track where it was, sometimes, destroyed or could cause other forms of major malfunction.
Dealing with products of varying sizes, sensors to detect the successful retrieval of a product seem like they'd be necessary no matter how things are setup, but by designing it with the robots on top, a failed grab, at worst, results in the product remaining somewhere near where it is, or in the wrong bin -- but not stuck in the way of the track.
Coupled with sensors to detect "when the product was 'lost'" or where the product was failed to be retrieved from, the system could attempt to "retry the operation", avoiding operational shut-down/having to rely on the skill of operational staff to identify a failure condition early enough to prevent problems.[1]
[0] I think, not positive
[1] There's a study out there regarding factory automation that appeared shortly after Elon Musk gave up on "fully automated production" that talks about how bad people are at responding to "rare conditions" in an automated factory, resulting in costs that often outweigh the benefits of removing people from the process -- this sort of design appears to try to address some of the major things.
Worth mentioning: I never worked with the mainframe in my previous job but was located in the datacenter. Most of my knowledge came from asking why the suite had 20 or so "big red stop buttons".
Apparently someone was hospitalized shortly after it was installed. And like all things, it was replaced a year later by a much more compact device that stored somewhere on the order of 100 times the data about a year after I started.
The tape system was really interesting to me during the short time that it was there (upgraded shortly after I started); I didn't expect to be able to find any useful information on it but surprisingly, an old IBM Promo video exists on YouTube.
If you're interested in seeing what it looks like, I put it at 1:04, since that's where the 3D rendered exterior view starts. They can be extended and ours was much longer than that one in the video if memory serves (one end to the other of a complete DC floor diagonally; it forced a really odd layout for the original mainframe, cabinets, and then ultimately the tape library that replaced it).
I wanted to include a link directly to it on the last post but couldn't find it after a few searches. I had originally read about it in the context of Tesla's automation failures (or, possibly, some of the issues around auto-pilot). It referenced a post-mortem style paper analysing the the failures and where GM failed to realize that factory automation is a "hard problem" back in the 80s.
I wish this video was longer and went into more detail, although I guess Ocado like to keep things light as they make money on selling this tech to other supermarket chains.
I used to be an Ocado customer for about 9 years or so and they were very, very reliable in terms of deliveries and order fulfillment (I rarely ever got substitutes) - I'm wondering if this tech and other things in the warehouse played a huge role in making sure inventory/supply always met demand.
When doing a similar online shop with ASDA we get substitutes/missing items/unavailable items every week, and the vans are always late.
Yesterday our large Waitrose order got delivered to our neighbour who has 24 hour carers on rotation. Turns out they took it in without questioning it. We complained our order hadn't arrived and they re-picked it and delivered a few hours later. So the last mile (or 9 in this case) is also important!
It's only a matter of time before the products are packaged to accommodate automatic systems picking orders; at the moment (as the video shows) the robot has to be adjusted specifically for the wide variety of packages.
Anyway I was thinking, why not just have a long row (can go vertical as well) of product dispensers where all orders just go under? I mean they'd pass thousands of dispensing stations, but the output would be really consistent. And there'd be shortcuts here and there. I probably have baggage handling in mind.
The problem that comes into my mind is that dispensers are tech, while bins are literally a bunch of plastic/cardboard/metal.
But on the pros, your system is a FIFO queue, while the system in the video is LIFO, the latter is not good for goods with an expiry date.
I am currently working in a warehouse with 80k unique items, i am planning to automate the system (because currently it's using little to none automation, mainly because there is a conflict of interest) and a system of bins with the pickers/refilles zooming on the top is the best solution I've found so far.
I think that i will make the project open hardware/software.
Food factories are FIFO, and you want FIFO dispensing, so I can only imagine the optimal solution is FIFO transport. I could readily imagine "magazines" for anything coming in cardboard boxes - spring or gravity fed, high-aspect ratio stacks in some kind of frame. Specially built container trucks which are easy to load with these "food mags". 6Axis robots to load them into the factory.
Spent mags get put into an empty trailer and sent back to the factory.
Loose bag items are trickier, but if you put those in reusable plastic totes, and put those in the mags, and automatically eject them into another mag, I think it would work well with the existing suction arm ones.
"We'll use the huge amounts of data that we gather to understand what customers are most likely to order"
Were regular Ocado users and its been a life saver during the pandemic - but - we have noticed that we keep ordering the same produce each week, the amount of new items is minimal compared to when we used to go into the store. Yes they suggest things but the suggestions (for us) have been poor whereas in store we would tend to try more new things. This may just be us and the algo works better for others but I think it should there is still much room for improvement in recommendation algos and there is a discovery challenge with online vs offline for groceries.
I was thinking the other day, if stores are shutting down in San Francisco to avoid shoplifters, will they switch to purely online ordering, or like some existing stores where you order from a front desk and the picking/packing is done by warehouse staff. I remember using a department store about 8 years ago in Edinburgh that operated like that.
And if humans aren't walking the aisles, will we see a change in product packaging to focus on shipping and picking/packing efficiencies rather than shelf appeal?
It's also how every supermarket in the neighbourhood in Barcelona I lived in worked. You could either buy pre-packaged produce (like two apples in a styrofoam based covered in plastic) or go to the produce section and point at what you wanted and an attendant would place it in a bag, weigh it and put a sticker on it which would scan at the register.
Another thing I hated about many European supermarkets were the one-way gates at the entrance (like subway turnstiles) and the only exists through the registers. Walmart started doing similar during Covid but at least you could exit through self-checkout.
I think it probably was - rings a bell. We were buying a child's care seat. I remember going there, flicking through a catalogue in a waiting room, placing an order and paying at the counter, and then waiting until it was brought to us. Seemed novel at the time, to us Australians.
Major retailers here have click-and-collect but I don't think it was common 8 years ago. A big hardware store like Bunnings only started selling online in 2018, I think.
I wonder if this means we could reduce the amount of packaging and ink used. And just stuff products in brown boxes without the need for reflective inks and plastic viewing window.
I suspect it's more efficient. The Amazon video mentioned that they found randomizing bin locations to be more efficient than any grouping. That's not particularly surprising when one considers that many non-physical storage algorithms benefit from randomization, but it's kind of fascinating that the physical storage of goods is a chaotic mess apart from the database that maps products to bins.
That big area where it is just stacks of bins and the robots that move them around would be a great place to set a chase/fight scene in an action movie.
I wonder what happens when a robot gets physically stuck in the middle of that grid. Do they have a trolley that humans can use to skate over and dislodge the offending device?
Or they have a boss-level robot with a bit of rubber on the wheels and you don't even have to fail in neutral.
I kind of wish this would turn into a giant robot wars for a few days of the year, maybe you can buy a ticket to control one of them for the day, if only it weren't such a waste ^^'
Ocado also bought kindred last year. I was looking at cutting edge robotics companies that might have rocket ship growth, and kindred was on my radar. The sum for the acquisition seemed on the low side. Robotics is generally hard at the moment, not sure who will be the Google-equivalent.
That seems like stocking bins is a solvable automation problem too. Trucks can be unloaded by vaguely smart forklift bots. Box moving bots are a thing already. Box cutting bot? They showed an item picking bot in the video: look at the barcode, maybe do a little tetris and fill a bin with them. Empty box crushing bot? Repeat.
Pallets shift in trucks, often requiring the warehouse crew to climb in and untangle things.
Also trucks are often dropping off a few pallets and not unloading entirely. The warehouse crew talk with the driver, sign way bills, and other things.
Yes these things could be solved but not by one company or quickly.
I always thought that most shops could/should have been replaced by giant vending machines long ago. Just be sealed-off warehouses with multiple dispensing units on the public-facing side.
Maybe someone like Amazon will try something like that.
This sounds like the future for high crime areas. Online ordering or vending style storefronts where you pick on a screen or with your phone and it drops a bag out the front with your items.
> Although 5 hours is quite a bit longer than what I expected.
I thought the opposite - even in a warehouse that huge, there are items where they only keep five hours worth of stock? To me that seems incredibly lean.
Of course, that probably only applies to really short life products, like packaged sushi.
I think the point wasn't that there are some products being restocked every five hours, and instead that the turnaround time from a truck turning up from a supplier to the products on that truck being available for delivery is five hours.
I do wonder what the environmental impact of doing it like this is. The embodied energy and materials used for all these structures and robots, as well as the energy to keep them running.
My immediate guess is that it adds a significant overhead to the products, but I really don't know.
I don't know either, but I expect the opposite. It seems incredibly efficient for what it does. I don't know how you could do the same thing with less energy? Obviously, it is possibe to optimize something here and there, but if we compare with humans doing the same job, and the energy they would need, this must surely be less...?
> Obviously, it is possibe to optimize something here and there, but if we compare with humans doing the same job, and the energy they would need, this must surely be less...?
The issue here is that the human needs the same amount of energy no matter what he's tasked with. Ideally robots will mean humans will be spending their time on fewer menial tasks, but the environmental impact of a living human is the same no matter if they're stocking shelves or doing any other job.
Robots might be able to give some environmental improvements in the workplace though. The areas they inhabit don't need to be well lit or heated/cooled for human comfort, but that's only true for places and times where humans can be removed entirely. Even in a fully automated warehouse the moment a repair person comes in, or a janitor, or a exterminator, or any human employee you have to accommodate their safety and comfort.
I suspect the additional efficiency of machines will just make it faster and easier for people to consume irresponsibility produced and packaged goods
I don't actually know! But there are a lot of factors. One is the movement of the robots, another is their construction and the supporting structure, as well as all the electronics involved. Making electronics is very resource-intensive.
I often wonder about how to make these kinds of comparisons, but it always involves so many guesstimates that it quickly falls apart.
I think you should compare it to what's normally being done for groceries. You would get a trailer of a good. Then in your warehouse split that into smaller pallets, going with other goods on a new trailer to a store. There, a worker would have to put the goods on the shelves, and a customer later pick it into their cart. The store also uses lots of resources. Same with every shopper in a neighborhood driving their own car to the store, vs a van delivering to multiple homes.
Oh, I understand that. It's still a lot of robots, but maybe it means less lorries and, ultimately, a whole layer of small shops and maybe secondary warehouses that aren't needed. Maybe it does add up to less fuel and less steel.
It also leads to a lot more centralization, but that's a different discussion, of course.
If you think robots consume much energy and are bad for the environment, try switching them out for humans and see how much energy you need to keep them running.
Oh, I'm sure it's cheaper to operate, otherwise they wouldn't do it. But I am of the opinion that energy and many natural resources are extremely underpriced when one factors in climate change and other externalities like the impact of exported e-waste.
Would you need to compare it to the environmental impact of the lifestyles of the dozens / hundreds of human workers who would be staffing a traditional warehouse?
I don't have a source handy but I remember reading something a while ago that said the carbon footprint for someone to drive to the grocery store and back is larger that the carbon footprint of all the other transportation involved in that product.
From a simple mass ratio perspective, these robots and a human probably carry about the same amount of stuff. Moving a few kg of robot units versus 80 kg of standard white male units, I can see the efficiency gain there. That is of course just one aspect, but I think you're being a bit quick to jump to guessing about conclusions of it being 'significantly' worse.
It's probably non-trivial, but human time also has a high environmental impact, because we eat and use stuff (think about a parking lot full of shoppers or employee cars). If we are going to pay our environmental human cost anyway, better if our time is not spent doing boring things that we do just because of the money.
I’m sure QuickSort is being used somewhere in there…
…though I assume you’re asking specifically about the quasi-autonomous movement of the tote-bots? There is no one single “algorithm” to represent or model their movement, it’s a complex control system.
Probably various optimization algorithms. You'd want to have your robots move the least distance for the work required since that minimises energy use and wear and tear. Some form of knapsack problem would needed to be solved for packaging the goods for delivery optimally. Then there would be scheduling needed to handle incoming jobs effectively.
It's not a difficult problem space. You can use quite simple algorithms for planning the movements of these robots. And you can improve them later without them affecting the rest of the system, which makes this easier than most IT systems which have all kinds of hairy dependencies.
If two grocery carts crash into each other it's not the end of the world. In contrast, if you're writing a distributed database system and you corrupt data then that can be very serious.
Uhh.... Hairy dependencies are very much a problem in this space just like any other type of IT system (in fact worse cause the code is usually written in C++ - so no package manager). Plus the tooling is non-existant... You will still need some type of client server architecture/failover redundancy. You also don't have very many options when it comes to testing. Writing simulations for this and running tests is very non-trivial. The algorithms themselves will need to take into account failures. What happens when one of the robots fails/wears out? (and this WILL happen) Deconflicting the robots is certainly nontrivial. For efficient solutions in this field theres a whole bunch of fairly non-trivial algorithms in Operations Research and planning that exist.
Things like controlling the robot arm that takes individual items from the boxes require state of art computer vision/AI techniques and can't really be done with classical approaches.
Optimization of the robot movement across the grid might be done in various ways with different tradeoffs, but in general the study of algorithms for multi-stage planning is also a subfield of AI.
but A* is not AI. Routing is like the easiest part and we figured that out for games and other applications decades ago. Use a 3d data structure where the third dimension is time, not exploring branches that cross a cell during a time where it is already marked as occupied. At least that's what I did for automating the air traffic control game from the bsdgames package on debian (/usr/games/atc).
Either way, GrumpyNl asks a fair question. Some parts, sure: machine learning aka "AI" for working with this vision system is probably a lot quicker than manually figuring out rules for recognizing items. But the amount of emphasis they place on it, that this is all run by one big AI? It sounds to me more like they mean "AI" in the same way that 90s games said "CPU" or "AI" or "Computer" player, when it had nothing to do with machine learning.
Couple years ago when people started doing machine learning for applications like computer vision in more visible, consumer-facing products and then marketing started calling it AI as well as, gradually, any software program that does something useful or seemingly clever with some degree of automation?
Technically interesting, but I hope it never replaces my local grocery store. It gives me kind of a dystopian feeling of a world where people are living a completely separated life, only interacting with machines. That will not lead to happier people, nor to more fulfilling jobs.
Most people working at such mundane repetitive jobs already behave and feel like machines anyway. Do you really think they feel happy and fulfilled, or even see a way out?
You are right about the majority of those jobs, I guess. I maybe shouldn't have brought in the quality of the job here.
I actually wanted to make a point against reducing social interactions in everyday life, especially with people outside of our bubble.
But well, I don't really see a way out, that wouldn't be completely illusive at this point, like, valuing well being of humans and nature over productivity gains and because-we-can in technology. That doesn't justify worsen the situation though.
As for me, I try to go only shopping in independent and ecological grocery shops, because of the products and because of the more welcoming atmosphere. Working there is not so bad, it's actually quite nice mostly (I can tell from experience).
Yes I prefer some underpaid human to wait on me too. But I see the world going in the automation direction at breakneck speed. If we can figure out where people fit in this brave new world (i.e. UBI) then its a utopian future, not a dystopian one.
There's no way this is profitable. Local supermarket is similarly sized and there are maybe 30-50 employees visible in store, each payed $3/hour. This has hundreds or maybe thousands of robots, even if only 10 robots broke down every day it would be more expensive than paying people. Even if I completely ignore initial purchase costs, I just don't see a way that after 1-2 years this isn't massive money drain.
No way? That's an overstatement. What if the real estate is a lot cheaper because it's well outside the city in an industrial zone?
And what if this operation is large and efficient enough to service a population x10 the size of a similarly sized supermarket with just a handful of technicians (and an army of delivery people...).
A million items a day? It all depends on margins and overhead etc, which we would have trouble estimating here. And even if its not "bricks-and-mortar profitable" it may be "wall-street profitable"
Also interesting to see the robot arm being used to move the products that currently come in boxes (suitable for humans to unpack) if robots become the norm will we see them being distributed like electronic components on reels/tape that are easier to unpack mechanically
https://www.google.com/search?q=reels+of+electronic+componen...