Reducing Industrial Packaging Waste With ($200) AI Tech

If you've ever received a small package in a ridiculously large box, you may have wondered 'what on Earth is going on?'. It's not just wasteful in terms of material resources of the planet, but also the costs of shipping must mean the company is also losing money? Given that around 90million parcels are shipped every weekday in the USA, this seems a crazy amount of waste. I set about trying to find out...

In doing so, I pitched a project idea to RS and NVIDIA to try to explore the problem, from both the human side and the technical side. I recruited Interaction Designer, Brian Schwab, who I knew from my time at LEGO to help with this tricky problem! We used NVIDIA's ~$200 Jetson Nano Orin computer to do the clever AI tessellation, or as I called it, to 'Tetris' all the items to be shipped into the smallest possible box. We also coded it to use the least possible amount of AI 'compute', and instead used spatial heuristics to simplify the task rather than burn-through CO2. More on that later.

The exciting thing for the Instructables community is, this is something we did as a pilot for a multinational company, but the code is Open Source on GitHub, and you can use it for whatever you want - whether you're packaging your suitcase for your holiday, or perhaps more likely running a small business, and would love to reduce packaging waste. Given the diverse bunch of enthusiasts through to professionals here, perhaps some of you even work at some of the big-box distributors, in which case, please get inspired, and check out the more detailed 12,000 word industrial series over on DesignSpark.

Fundamentally, this is a project about using 'tech with good intentions', as per the pending Contest - it is a way to Dream A Better World, but it also depends on folks sharing this with people who work in packaging design, logistics, sales and marketing, as the problem of wasteful packaging is not just about 'bad boxes'. The use of AI is, as we have said, as minimal/responsible - and indeed, because we are not relying on Cloud compute, it's less energy intensive, and also more private (a big concern for anyone, not just big companies, these days!).

We've worked with the operations team at RS, and really tried to discuss the thorny issues of AI in the workplace. As you'd expect this makes many folks apprehensive, but we hope you'll see from the videos, and previous work (DS) in this space, the key seems to be to involve people at all levels, as early in the design process as possible, and continually throughout as well. Many thanks to the staff for taking a leap of faith to explain their preferences of what aspects of the job it's beneficial to them to have AI support.

At the time of writing, new EU packaging legislation (PPWR) is due to be in effect in 2030, and is highly likely to significantly influence the USA and global shipping as a result - in short, if you ship more than 50% 'void' or 'air', this will be subject to a penalty fine, which is not trivial. So many products will likely be redesigned and repackaged in consideration of this taking effect. So I'd love to hear how you get on, and do leave comments or contact me for any follow-ups. Enjoy the project, and best of luck with your own more sustainable design projects.

Disclaimer: This project has been shared in good faith to promote more sustainable use to AI, to reduce packaging waste. However, although every effort has been taken to ensure this is robust code, we cannot accept any responsibility for any issues arising from using it. You do so at your own risk.

If you also like to listen to a Podcast about this project, check out my interview with Talking Rubbish Podcast, Ep 95.

As mentioned, this is quite remarkable to consider how the Jetson Nano Orin, (~$200) by NVIDIA is an incredibly powerful microcomputer that allows us to perform the 'AI Tetris' operation on various packages.

You will need to download the GitHub files here: https://github.com/DesignSparkRS/boxing-match

As well as follow the general installation instructions on Jetson Jetpack 6.2 (at time of writing), which requires some careful installation, and a bit of concentration, but has worked for others who have tried it!

Although instructables guides are usually about finite/explicit steps, this is more holistic in that we're documenting the process, so you can benefit from the learnings as we worked through the project, as of course your company will differ from RS's distribution, so you will have unique challenges with the products you ship. One thing to note is that if you ship 'mostly random' product collections, like RS, this can be configured to be a 'general' or 'best fit' model, with looser tolerances. However, if you run a company shipping the same thing every day, then you can get very specific, and have very high tolerances. On this latter point, we've added a examples to demonstrate how RS can do this with a 'kit' with 13 specific items that do not change. Furthermore, if you feel that it warrants it, you can move from a 'off-the-shelf' or 'generic' box, to ordering a bespoke one, which may cost more in unit price, but save price in reducing waste. So as you can see, the specifics are down to you.

You will also need some basic computer peripherals, monitor, keyboard, etc. But really these are trivial details, the main crux of this it to modify the input data to suit you own packaging setup. See GitHub.

-

Why No Camera?

We did initially use a webcam as we were looking at image recognition and scanning in 3D, but interestingly, we realised this was 'overkill' and not actually needed for doing a vast amount of the tasks. This simplified the project dramatically, but if you wish, you can of course use one to collect the detailed 3D 'point cloud' of a product, rather than just the gross dimensions, but for RS, most of the time X, Y, Z dimensions were fine.

-

Would it run on Raspberry Pi AI Hat or Arduino Uni Q (or other AI enabled SBCs)?

It's not impossible, as the project is designed to run with minimal compute. However, the bigger your dataset (RS's Nuneaton Operation handles tens-of-thousands of boxes per day, from millions of stock items), so the Jetson Nano Orin has the power to cope, but it's possible you could 'port' this to other hardware/software, and hence the Open Source ideology behind this - thanks in advance anyone who takes this to new places!

For the clickbait/ragebait, I took a picture of me sitting in a massive box, which shipped the few items I could hold in my hands. This is not to single-out this company, but the fact is, this genuinely happened, and is not so uncommon people don't post about it online. You'd expect such things to be covered by liberal-leaning papers, but if even tabloids like The Sun feel compelled to cover the story, it seems it's baffling to most people.

However, being an designer/engineer, part of your job is to 'dig deeper', and not just point out what is broken, but to try and understand why it happens. One has to assume there might be a problematic, yet prosaic trade-off where sure, the cost wasted in things like this is eclipsed by the cost of the intervention to correct it. Or to put it more explicitly, if it costs more to have 10x more box sizes (so one can select the best one) this cost is greater than the extra postage and packaging material costs, so the action is to leave 'imperfect' as is.

This may sound like that's the end of the matter, and a resigned *shrug* is all one can do, but firstly public pressure is making a dent in corporate social responsibility - as I should say to RS's credit, this [DesignSpark innovation] project was approved and began before we had any knowledge of the legislation coming down the pipeline! Indeed, 2 months into the project we met with the Head of Packaging at RS, Liam Dowds, who was only too happy to engage, as he was aware of the PPWR legislation, as mentioned above. This combined with Extended Produce Responsibility (EPR), which means if a company produces a thing, like packaging, it has degree of responsibility to that things impact in the world, so in other words it *pays* for putting excess packaging waste into the world.

In an ideal world, all boxes would be shaped to fit the size of the content, but this is actually inefficient in its own way, as it leads to complexity, which is wasteful in other forms or staffing, transport, electricity, computation, inventory, etc. so it's not as simple as 'just make more sizes of box' - the trick is to find optimised 'sweet spots' where things statistically cluster. The other side is to explore how to fit more stuff in a given box with less void/air - and hence the Tetris approach was compelling.

This is the illustration from the initial pitch document I created. As you can see the general idea persisted of 'Tetrising' (this is an American website, so I can verb nouns in ways that would get me imprisoned in the UK), so I will enjoy the liberty here.

Anyway, another interesting point was that I initially assumed that to fit things carefully into the box with as little room as possible to spare/waste, would hinge on 3D scanning the items individually, so we could for example put a bag of screws in a 'hole' in a roll of Duct Tape, for example. It turns out that, as with much in life, and as described by The Pareto Rule (or 80/20 Rule), that this would require a massive mount of work to scan things (we're talking hundreds of millions of items potentially), that instead, simply using the X, Y, Z dimensions would get us to a pretty good result - at a faction of the cost and complexity.

And for those of you worried, as I am too, about AI and the environmental cost, it's actually not a 'win' to compute every nook and cranny of a screwdriver, for general packaging - it's more efficient to treat it as a 'primitive shape', like a cylinder or long thin 'box'. So this is why our system is not needing a Quantum Computer, but a $200 Jetson Nano Orin.

I has this with wonderful '20/20' hindsight, but as you might guess, this took us a few weeks to figure out, as when working in tech, there is a very real tendency to 'geek out' or consider 'tech for tech's sake', even without realising it. So it pays to 'zoom out' and check one is really doing things in the most effective way (not necessarily 'cool' techy way). The goal was to reduce waste, reduce energy, reduce complexity - so X,Y,Z dimensions (a couple bytes) are more streamlined than a 3Megabyte file for a 3D point cloud for a tools which is better modelled as a 'box', 99% of the time.

The exception, as we will see later, is if you plan to ship something *repeatedly* and in which case, this would justify the 'one-off' super-accurate method. But for shipping thousdands of items a day at random, this honestly seems the smartest choice.

A lot of what I love about being a designer/engineer - is meeting people from different walks of life, and understanding their jobs, culture and discovering insights which end up being invaluable later down the line.

The team at RS Nuneaton were incredibly helpful and very open to what aspects of their job were straightforward and what were tricky. Some of these points are best left to humans rather than robots or AI, as they are highly nuanced, and are infrequent (meaning the AI can't get much data to 'train' on what is an every-changing task). However, some tasks were already automated (even without modern AI, I should add), and some were in that grey area of discussion if AI would help or hinder...

It was somewhat ironic that I ordered a Jetson Nano Orin from RS Online, and it was shipped in a box which was a bit too big, whilst working on this project. You can scoff, but actually this has become a perfect empathetic 'test case' as I've presented this work...

If you were a human packer, packaging hundreds or thousands of things a day, you have a few seconds to look at an item, and apply your 'best guess' as to which box you will put it in. You then have to make up that box, using parcel tape, etc. If you guess it too small, it's not the end of the world, but you wasted time. You're not allowed to chuck it on the floor for later (as this would be a trip hazard which would pile up fast!). So the only rational decision any sane person would do is to err on the side of caution, and get a bigger box, to be sure it fits. So I 100% defend the 'correct' decision of the packer here, for the avoidance of doubt.

The fact that it would actually fit in a smaller box (the next size down), with 1.5mm clearance, is not something I could 'blame' the person for. Remember these things are coming to us at random, this is not a task where one can 'dial-in' specifics, like say a Bakery. So it is unreasonable to expect a person to 'guess' this would fit in the smaller box, just so.

However, this is also where the AI can truly make a difference. Because the AI can check the gross dimensions of the product, and it knows the internal dimensions of the boxes, it can verify that the Jetson Nano Orin does indeed fit in a 'B2' (or whatever) box, and the human packer can build one in confidence, knowing it will fit for sure!

This is an empathetic example, which went down well with the team, as it shows how AI can help reduce waste, by performing a 'check' which would be excessive for a human to do. Of course you might be wondering 'but what about orders with multiple items [greater than one] - how does that work', and we'll come to that later, but for now this was a great answer to show how AI can be an 'assist' a 'Co-Pilot' as it were to help people working at speed, with little time to laboriously cross-check details. AI is good at data, humans are good at getting it done, and being aware of when the AI is wrong, and doing the holistic goal.

As mentioned, AI is great when there is a definitive answer like 'Can this [Jetson Nano Orin box, 230x185x75mm] fit in this box [B2, internal dims, 231.5x198x98mm]?' - the math is simple, and so the answer is unequivocally a 'yes'. The human can proceed in confidence.

However, 'Common Sense' is something we often feel AI lacks, and yet we often forget that a child will touch a dirty thing on the floor, only for parents to wrangle out out of their hands before they eat it! The baby has no awareness of 'why' something was a 'no' and it takes many years to give a holistic framing of why sure - don't eat stinky dog poop, but why not that brown stick, when brown bread sticks are okay, and so on. Then factor in a 1 year old has way more intelligence than any computer, and it still makes mistakes! The point being, AI has to be taught everything it needs to know, and then many exceptions.

It is obvious to a packer that they should not put a Hammer with Lightbulbs (for the record we tried this and RS packers never made a mistake). However, I managed to create an order for a hammer and delicate microspeakers - and RS packers did arguably get this wrong. They should ideally have been packed separately. To be fair, the speakers did work, with no issue, but as they are delicate by nature, it would have been wise not to put them with a hammer. Like the Jetson Nano Orin and the 'perfect fit' B2 Box, an AI can perform a 'look-up check' to ensure that the packer does not make this mistake, as frankly, unless you have a expert knowledge of microspeakers, most people wouldn't recognise what they are, let alone realise they are delicate.

Now assume this for millions of lines of stock, and in millions, no, billions of combinations which are not idea. This is not a task of 'AI' but is 'Common Sense' which needs to be coded as rules for the AI to work. We need to explicitly say to an AI that it should avoid shipping a thing that is sensitive to magnetic fields, with a big magnetic thing like a motor. This sounds like one could 'simply' get AI to read all the PDFed datas sheets and simply 'learn' it all, but it's never that simple, and this takes considerable time, but it's fair to say it will get there, but when people say 'the robot apocalypse is coming', I'd say, it's not coming as fast as you think, because like a baby learning to know what to eat, and what not to, it takes years of patient 'parenting' and 'teaching', and that is neither quick, nor easy, no cheap. But yes, in 10 years time will most companies have figured out a lot of this, sure, but a lot can happen in 10 years time...

Although this project is Open Source, it contains a small 'demo' example of some of RS's stock items, as of course we're not publishing their entire inventory in public (although the Jetson Nano Orin can 'handle' it).

The example above shows how I manually entered in a X, Y, Z dimensions for some products RS sells, see the blue columns. However, there are some things between accepting the over-simplification (and missed opportunity) of 'gross' dimensions and a over-blown 3D scan of every item...

...We realised that when packing things in boxes, the ability for good 'tessellation' was pretty high, as computationally, things fit next to each other fairly unambiguously. The AI side of things can apply what is referred to as 'NP Hard Bin Packaging', but as mentioned this is computationally intensive (the clue is in the 'Hard'). So we tried to again simplify this to be 'less perfect', or as I called it 'NP Medium-Hard'. This is still better than most humans, but it's also not striving for perfect - again, Pareto Principle in action.

If you work in AI, you'll know this term of 'Heuristics'. But not you perhaps don't. However, parents likely will have heard of it, as it comes up in your 'child development' discussions if your kid went to a kindergarden. Heuristics are the thing that happens when a baby is truing to put a square peg in a square hole, as shown, or indeed, failing to - it's the 'Trial and Error' part of how we learn. In AI, this is why terms like 'feedback' and 'reinforcement learning' come up also, and sound very much like teaching a baby, cos in many ways they are!

This is perhaps not the place to deep-dive into this, more on that here, but the gist is we wanted to use some heuristic methods to avoid 'brute force' computation, where you are just slavishly processing all options, even if much of it is silly. For example a baby might try the square peg in the round hole, but it doesn't try the cat in the round hole, so nor, computationally speaking, should we.

Okay. So where we got to was to do a little better than 'boxes', but not super-detailed scans...

We realised some things were just as simple as a box, cos they were a box, so we called these 'Rocks'.

We realised some things could be rolled-up in one dimension, as a secondary option, should this help the Tetris Algorithm deliver a batter result (ie smaller box), so we called these 'Rugs'.

Lastly, although a bag of screws will not behave like 'water', and flow into any space, as it is still bound by a bag, it nonetheless has a bit more flexibility and freedom than just a 'Rug' / rollable item, so we called it 'Rubble' to imply it can perhaps fold in 2 directions, and has a few basic alternatives.

These are not exhaustive and that's precisely the point. We made the spreadsheet have the basic gross dimensions (in blue columns) for when the task was straightforward (Jetson Nano Orin in a B2 box - alone). But we also created these heuristics (along with the 'do not pack X and Y together' rules) such that the algorithm has some latitude to try and few options, without being computationally excessive.

As mentioned, the above examples were for when the packer is confronted with thousands of unique orders. With very few exceptions, one would not expect the packer to 'retain' or 'learn' anything with this vast randomness, so this is where AI can be truly helpful and lighten the 'cognitive load' for the human.

The converse is when you have designed a standard kit or assortment of things which will be repeatedly shipped. I had a small project I was designing that used tools for graduates, and so this collection of 13 objects was going to be 'standardised' the same way you would a spanner set, and so dedicated attention could be paid to ensure that we optimised for as little space as possible - and utilised the Rock, Rug, Rubble heuristic.

Just to illustrate the point that there can be diminishing returns. This graduate project was looking at perhaps a small run on 100 units, as a pilot. So it's not really worth commissioning a bespoke box. Better to use a 'standard size' box RS already have in inventor, let's call it 'B2'.

The 'loss' here in a B2 box is 17.2%, which is better than 50% by law, and in terms of experience, does not really need any 'filler' or padding paper, as the collection of items is very close fitting and snug anyway.

If we allowed the AI to 'fully optimise' it can reduce it by a further 2.9% - but this would require RS to buy/source/design a specific box for a trivial an temporary order. So it's easy to see now, why, when looking back at the very start of this project asking 'are companies crazy for being this wasteful?!' - it's a bit more complicated than that, but of course this project works through the stages to how it might be a bit better through incremental improvement, that doesn't 'boil the oceans' in intensive brute-force AI computation!

Although our sponsor for the project was NVIDIA, it might sound 'risky' to try and use as little AI as possible, but this is of course a 'relative' statement. Using a Jetson Nano Orin ($200) to run a database from a global distributor, without resorting to cloud computing (a cost issue and privacy risk, for some), is actually a very frugal use of tech, when you consider it's impact for just one site is millions of boxes per year.

For me this is 'more responsible' use of AI, than just 'burning tokens', which at the time of wirting seems to be a growing backlash that there is almost a bravado of not 'how good is the work' but 'how many tokens did you burn through this week' [doing tasks which perhaps didn't need tokens in the first place].

When I worked at Dyson, one of the marketeers told me that a campaign that works well in pencil storyboards will usually be amazing once they add the colour and filming on top. But if the storyboards such, no amount of 'lipstick on a pig' will save a bad, confusing or off-brand idea. I feel the same is effectively true with AI - we need to design things to be as 'lightweight' as possible. Otherwise, we risk making the equivalent of Code-SUVs, sure they are comfy, but are you really driving a 2 Tonne EV to the mall for the a pint of milk?

On this note, I was aware that our AI 'Tetris' algorithm could be displayed on a monitor at a workstation in front of a packer at RS, and this would look all very cool and modern and sexy, (make us look cool)... but I was also not afraid from realising after coming through all of this discussion, thinking and prototyping, that possibly a version of the output could as simple as tweaking the packaging print-out, as shown. If this is what a packer needs as a 'headsup' that a given item can for sure fit in to a given box just so, then does it need a fancy jumbo-tron monitor, or will this do?

I met Brian Schwab a few years back via LEGO connections. I've worked at a bunch of really great companies, both an an employee and now as a freelancer. Whatever project you eventually define and set challenges to solve, you'll never be able to do it without someone who can do something you can't. I'm very grateful to Brian for joining the team and helping realise this idea from my goofy sketches at early pitches, to the delivery of the prototype. This is a continual theme of my work - coming up with an insight, pitching a direction, realising it'll morph and change along the way as more is understood, and learning (and failing) together, in order to arrive at a robust solution.

When working as an 'outsider', it's often a mixed blessing - as on the one hand, you are bestowed with a lack of baggage of being in a company as an employee. The counterpoint to that is you are not 'one of the crew' - so trust has to be built slowly, and with care. It take someone to 'bridge' both the internal and the external. Liam Dowds, RS's Head of Packaging, was that person - and although I can't say much, he is leading the future internal phases to see how to integrate the learning of this pilot.

I'm keen to not imply this is 'job done', it is most certainly a 'pilot' - small and provocatory for sure, but to properly scale with many stakeholders and more complex datapoints added will take time and patience, as well as considerably more resource, but it often starts with a 'spark'.

This is what I hope this project is for others - a Spark - that ignites something in the R&D, Innovation, Production, Sales, Marketing, C-Suite, etc. - to realise the story of sustainability need not just be one of reluctant coercion into legal compliance, but can genuinely be used to drive-through necessary updates, promote internal training, and improve job satisfaction overall.

There is a long way to go, but thanks to Liam, Brian and the Nuneaton team (you know how you are!), we look forward to advancing this project in UK, EU and globally in time. 2030 seems a long time away for some, but for many businesses, this is very close indeed. Find out more here.

Please take these figures as a 'straw man' - they will of course vary with your operation. However, the ratios (e.g. how much graduate : supervisor time) may provoke useful discussions to tweak to your preferences. More details can be found here, but I'm always reachable here if you fancy a chat...

For those of you wanting the more detailed cut of the video, hope you enjoy this also!

Aside from visiting the operation at Nuneaton, I also 'soaked up' various other sources, and found Eva Yin's exhibition at the Design Museum, London very handy for framing my thoughts...

Her work explored how AI can better tesselate fabric patterns onto fabric, taking into consideration the warp/weft of the fabric, to reduce waste. I figured if this can be done in 2D by a student, then one would hope with a bit more industrial / computational budget, a company like RS might be able to do similar for 3D problems of box packing.

It was not lost on me that casually going from 2D to 3D would not simply 50% more effort, but likely a cubic or quadratic law of complexity of increase! But as with so much in life, you kind a have to walk the line between total AI hype and BS, and selling the 'art of the possible'. I'd love to say that this works 100% of the time, because of course it does not for anyone in innovation that is truly taking chances (we actually started a prevcious project to do with LLMs which flopped before this in fact!), but with each project you learn more, invest in people, and that has a compounding effect on the capability of a team to know better what to do next time. I feel it's a rarely discussed thing in innovation - how much it's like a 'Phoenix' - good thing can rise from the utter failure of a previous idea. I think the trick is not to get reckless and 'bet the farm'... the previous project to this 'failed fast' after 2 months, and although it meant our budget was less for this, it was not all gone, so we 'licked our wounds', reset and make progress based on that hindsight - for the better I feel.

Liam would be surprised to learn that a packaging challenge started off from first working with a LLM to explore language processing between different languages, in Hong Kong and China.

Before that, this idea was first pitched by me to another team some 4 years prior to this, but then the NVIDIA Jetson range was not affordable and nor did it have the community and use cases to help us code it and get it working. Which is a long way to say 'ideas are one thing, timing and capability are quite another'.

If you work in innovation - be prepared to 'pause' and 'phoenix' a lot more work than most people's glitzy websites and resumes will admit.

Another enjoyable aspect of the project was to reach out to people like Henry Hargreaves, whose 'Crisp:Air ratio' project I found whilst doing general research for this. It got me thinking about what else might struggle with PPWR, and hence need a redesign...

I had a saucepan arrive, after I broke the handle on it, and it made me wonder if more saucepans in future might be 'DIY' at home, and you screw the handle on, as this would keep it under the 50% limit. Likewise, weird items like watering cans, which I got for my young son's gardening 'enterprise', and arrived very wastefully - but perhaps the spout will just 'press-on' in future?

After presenting this work at Birmingham's National Exhibition Centre, at the Packaging Expo in 2026, I subsequently posted on LinkedIn about how I thought the Penguin Biscuit was a great example of how something which was a 'byproduct' of the packaging process, but ostensibly 'waste' - the 'flap' or heat-seal - was used to create a unique and memorable experience, for free!

I can't say much, but this post ended up with me becoming a project consultant for an established brand to help with similar thinking on their product range.

So, I've been keeping a scrapbook / photos of things I think PPWR and EPR might 'be the death of' - and hence, like a Phoenix, we might re-birth a project and it's packaging in a new light, not just to be compliant with legislation, but to be in-keeping with a social consciousness of wanting to not be wasteful.

I don't know if this is a guaranteed way to 'dream a better future' [as per the contest here], but it seems that some things with be radical and some thing will seem tangential or even incidental, but one thing for sure is we cannot simply continue to 'chuck stuff in any old box' and not worry about it. Those days are over. There will be those who see it as purely a annoying bit of eco-legislation, and there will be those who think of it as a way to make things more compact, modular, interchangeable, repairable, and more...

Which will you be designing?

What would you buy?

-

If you'd like to chat more, please drop me a line here, on www.judepullen.com, or LinkedIn.