Hacker News Reader

Comment by jacquesm | Hacker News Reader

And mine is just as universal, and likely also not the first but the first one that got wide visibility.

This guy did a great job of addressing some of the shortcomings of my machine, specifically in the number of output bins (18 instead of 12), but it is a regression in other ways (built out of Lego so it will break frequently, mechanical part movement from the belts, which will lead to frequent jams and will wear as well as that it is slow and more timing sensitive).

All in all a nice effort and very nice video production but not nearly as much of an improvement that would make it reproducible/mass marketable.

To properly build a machine like this will take a significant budget and a lot of hours behind cad software to make it out of industrial parts rather than Lego. I you want to make something mass marketable then you have a very long road ahead of you.

View on HN

I believe what he means by "universal" is the ability for the machine to recognize parts it has never physically encountered before (since it's trained on 3d models).

It seems like increasing the number output bins shouldn't really matter much, since with either machine you could just reprogram it and run parts through multiple times, right?

You can train it on three D models but that's just a stand-in for the parts as though you have them. But - and this is quite important - if you want to achieve some efficiency with the 38000 or so parts that Lego has made over the years then you're going to have to cut a couple of corners and one of those is to give some more weight to the parts that you do encounter and to train your model to the point that it will come up with the right category even if it has never seen the part at all, virtually or IRL.

That's roughly what I got it to with showing it samples of wheels but not all wheels and it detected wheels that it had never seen before just fine and so on. This is a lot easier than it might seem, you'll get a good match on a class but no good hit for any of the subclasses: that's a new part right there.

And yes, that's exactly how I did it, multiple passes. But more bins is certainly useful, I just ran out of room, the machine was already getting quite large, and the belt and frame I used was fixed in size and dictated a lot of the dimensions of the machine (it was a re-purposed running trainer).

Hey, Daniel West here, the creator.

First of all, as I mention in the video, this project is inspired by the machines built by akiyuky and jacquesm@. So I'm certainly not claiming this is the world's first LEGO sorting machine!

In the video I try and focus on the unique aspect of this machine, which is the Universal part. Because the CNN is trained on synthetic data, the machine is able to recognize any LEGO part that has ever been produced. This is with the notable exception of ignoring printed parts, mold variations, and oversized parts, as well as parts that are not present in the LDraw part library. These restrictions bring the tens of thousands of part categories down to just under 3000. Accurately classifying 3000 categories is prohibitively difficult using manual labelling alone, simply because it's very hard to get physical access to all these different parts (not to mention the labelling process itself is quite laborious).

Is it overkill? Totally. But one goal of this project is to be a testbed for the use of synthetic data in a real-world classification system, and I'm quite happy with how well it ended up working.

I have produced another video explaining the AI system in detail: https://youtu.be/-UGl0ZOCgwQ

It has also been noted that the machine would not be mass marketable in its current state. That is totally accurate! This machine should be considered a proof-of-concept/prototype and I have no plans to try and create a mass-produced version.

Very neat project, congratulations on seeing it through. Note that being able to sort down to single parts is less useful than it may seem simply because of the number of passes required and that you'll likely reach your 'universal' stage a lot earlier simply because neural nets are good at classifying stuff they've never seen before as long as enough of the bits and pieces match.

How long does it run between fixes? How many kg of lego have you put through it?

Beware of lots people that will bug you to give them all your work, to this day not a week goes by that someone contacts me with a thinly veiled request to get a copy of the hard parts 'just for their personal curiosity' and then when you search for a bit on Google you'll find them associated with large parts resellers.

Highly annoying.

Best of luck with your project, curious about what you'll do next with it.

Thank you Jacques! Indeed, for practical sorting purposes going down to individual parts is less useful. But there are intriguing possibilities, like for example using it to keep track of your collection, or extracting sets from a collection.

The machine is quite robust to jams so long as you're careful not to insert oversized parts. That being said, I'd say there's a problem every 30-60 minutes, so certainly not something you could leave overnight. Running for hours and hours means that it will eventually grind itself into dust anyway (I've already burned out a motor!)

I'll probably end up sharing everything one way or another. I'm sure you'll agree that any real commercial application of something like this is pretty far-fetched so I'm more than happy to just treat this as a contribution to the body of domain randomization application research.

I guess someone could consider made-from-Lego to be part of the fun rather than a weakness.

It is odd to claim a first whilst also crediting you. He seemed to emphasize the universality in several places (I guess you didn't explicitly state this?).

I thought the photos of the parts bins spoke for themselves.

At some point the guy that bought Bricklink wanted to buy out the project and I told him to f-off because of the way Bricklink has been run, no surprise to see him sell it, likely he got many times over what he paid for it.

I'm happy to see someone continue with the project though, for me the fun was over as soon as the software worked and it turned into an engineering/business model affair. Bringing this to market is a lot harder than making a prototype.

> Bringing this to market is a lot harder than making a prototype.

Not realizing this for almost any product is one of the biggest mistakes technical founders make.

https://xkcd.com/678/ 5 years

I realize this doesn't add to the conversation, but I love the fact that you can mention someone on HN and they might show up for the discussion.

Reply on news.ycombinator.com