0:01 Okay, good afternoon. I am going to describe it, I won’t fill the full 20 minutes, probably going to try to describe a couple learnings from in practice applying deep learning machine learning techniques to cam in a couple different products over over the last, over the last decade or so. Maybe just briefly describe my background, I used to be a professor in machine learning a long time ago, then I created a company called circuit IO, which was acquired by Autodesk, which today is stinker cat, it’s probably the most used cat product in the world. I was part of Pier n, like many people here, I managed their Advanced manufacturing portfolio, then I became the CEO and co-founder of Octon, where we really created a like an integrated production platform that really enabled AI to automate a whole bunch of cam workflows. Managed software at 3D systems for a while, but then I was able to spin out two of the software entities that were part of Octon in the past, Arc and okam, we will describe them in a bit more detail, and I’m current ly the advisor, hacker in Residence for for these companies.

1:34 Just briefly, for the designer crowd, like a cam, you have a part, then you go into a separate program that turns the geometry into a program that then actually allows a machine to produce the part. Sounds very simple, but then there’s a whole bunch of decisions you need to make about positioning, orientation, fixturing, like nesting, process parameters, like in the cam product, but this depends on okay what’s the order State, what what other things have been ordered, depends on stock Materials, stock material availability, depends on which tools are loaded in the machine, depends on like like how would you work hold it, like geometrical features, which Machining templates, like it’s like a giant complex problem. So people think ah cam, it’s just like a print button, no, cam is really quite complicated, and that whole industry is a bit at a at an inass.

2:32 So it’s estimated to be about two and a half billion dollar software Market, however a lot of the batch sizes are becoming smaller, so where things were produced by the millions in the past, now often batch sizes are shrinking, because time to Market is reducing, lead times are getting shorter, so there’s the engineering effort in generating the program, can’t be spread out across many many parts that are produced. Also manufacturing processes are getting significantly more complex, like a three-axis milling machine versus a like multi multi five AIS Milling station, like it’s very very different. The, for those that haven’t seen like real world cam software, it has 100 buttons and pages and pages of things you need to be setting up. This is typically something a cam programmer takes years and years and years to create an intuition on what all these settings actually mean when you physically produce the part.

3:42 But yeah, cam programmers is a bit of a Dying Breed, it’s like if you look at the average, like the median age for like like milling machine programmers, it’s like 54 and increasing, so it’s like it’s a becoming a real problem, who is the next generation of people that are actually going to drive all these machines. And like it’s all the processes out there, Mill, like Milling, turning, EDM, cutting, bending, additive, composits, robotics, like all use software like this. I think it’s a real opportunity to try to make this much more accessible.

4:18 And so one of the companies I’m involved in, it’s called aranc, and they really try to automate the programming of welding robots. So you might, like, there’s a significant shortage of welders, and one way to solve it is to actually install welding robots, but in the past if you would have to generate a program to let’s say weld a trailer truck, a trained cam programmer would be spending a week to generate a program, very very labor intensive. Here you see the software that we developed, so this is browser based, you upload a cat file, we actually use, yeah, deep learning models, search, all kinds of algorithms under the hood, to try to automatically identify weld positions, to automatically identify how these should be merged or split, to automatically generate fully Collision free robot programs, where in the past somebody that was a cam programmer for welding needed to be a welding expert and needed to be a robotics expert, and that’s like a very difficult skill set to to to get overtime.

5:37 Here you see the the software in action, where we literally were able to go from a week programming time down to like an hour. This, for these companies, significantly changes the way they can think of a robot. In the past they thought, ah, we can use a welding robot when you produce things by the by the hundreds, now if you produce a single part you could even already start using a robot, so that’s what that software enables. And customers that have adopted this already really say okay yeah this is like game changer for them, from multi-day programming to just like hours, really changing the economics of what the robot can do, but also completely changing the Persona that can actually control the robot. Now they can just give this software to welder and the welder can start using the robot, which in the past really was not possible.

6:30 So that’s one company I’m involved with, we learned a lot of things here of how to use AI. Another company I’m involved with is called okam, very different, focused on the dental industry. So this is metal and polymer 3D printing as well as zonia, like primarily zonia Machining, where again in Dental Labs these people are not highly technical, they’re Artisans, they started adopting these Technologies because it really changed their workflows, but cam for them is a real issue, they’re producing often thousands of parts in a lab and these things need to be shipped out tomorrow, so like lead times are insanely short, and they have a hard time hiring and retaining these programmers. And so what we were able to do, again, create a cloud-based product that integrates with your existing Erp systems, they just pull in all the data and try to assist and automate as much as possible the preparation of of all these parts that are all unique, so this is batch size one for all of them, across different Technologies it’s always the same user experience. Again, here customers are quite pleased with the results, not only being being able to put more parts on a build, but also significantly reducing the cam processing time.

8:01 So from these companies I’ve learned a couple things the hard way, some things might be obvious, some things actually were for us a bit of an aha moment, but I’ll try to share three main learnings. The user experience for automation is super important, and there’s very different approaches to look at Automation, and then it already came up before, like what about data. So first automation, so first the user experience, most of the cam tools out there are a bit like dslrs, they’re like these amazing machines, I really want this, but then you need to get a course, okay, how do I use it, first time you take a picture in dlsr, like it gets really really crappy, and that, like, you need to have training and you need to like study the manual to to be able to get everything out of it. This was designed for an expert, and you need to become an expert before you can get most of the benefit from it.

9:09 If you compare this to like an iPhone, even with the first generation iPhone, the photo app looked exactly the same, you see what you get, and there’s just like one button that says take a picture. In the first iPhone the quality wasn’t great, but at least it was able to consistently take pictures, and over time they started layering AI, they started giving depth of field, they, like, now iPhones have basically eaten D dslr’s lunch by consistently sticking to the same super easy user interface, making everyone a professional photographer, and then layering more and more advanced functionality under the hood. There’s, there is still control on an iPhone, but it’s like hidden away, it’s not like all in your face the whole time, if you need to go in and and change the the the the point where it focuses, yeah, you can tap and then like the user interface appears, like this Progressive appearance of control, very very powerful. I think this is a great way that we discovered, like, okay, this is how user experiences should be designed when you’re, when you’re looking at a at an automated collaborative workflow.

10:30 Okay, another big one. When we were doing this work on on welding robots, a lot of people came to the software thinking that they need to be worried about the robot, oh yeah we need to like, what about the joint positions, and and what about being able to decide which like wrist position it it should preferentially use, like none of that really matters, like in the end the only thing people care about is where is the weld being laid down and what are the weld parameters. And so we, we have to completely rethink the user interface to be about the process instead about the machine, most cam software is about the machine, where here we really put the emphasis on on the things people care about, where are the welds being put down and what are the process parameters that you use for the welding. So we really needed to like flip around almost how the whole user interface worked, actually the first version of the product was focusing on the robot, and we had like rebuild the product to really focus on the weth, and then suddenly it started taking off.

11:35 A third learning on the user experience for automation is that although you might really hope that your software is 100% automated, it never is 100% automated, and the worst thing, especially in a production environment, if if people get stuck, the thing needs to ship tomorrow, and they let the automation run, and then for this one part, like they’re not happy with the decision. So although you think you’re building a highly automated product, what you actually need to build is first you need to build a completely manual thing, because people still need to go in and be able to take over, make manual decisions, I see some heads nodding, and then the automation can be added, you you lead with the automation, but people always need to be able to fall back to manual workflows. Again, a learning after the fact, it’s like super obvious, like we learned the hard way, we had a lot of frustrated customers in the beginning.

Good at like a a fourth point, and I, I’m not working at AO disk anymore, but like one of the things that was like really amazing about Tinker cat, Tinker cat is really Google Docs, like it’s has the most advanced data model for cat in the world, it’s fully collaborative, you can literally have multiple users all around the world collaborating on the cat file. On shape has some, already has very fine grained like Version Control and allows some collaboration, but it’s not at this level of collaboration. Having super fine grained like capturing of changes is key to get automation to work, it’s a bit how theado presentation also described, okay like this version control is key, when when you have one person is on the main branch and somebody’s making this feature change, that’s what AI is doing, AI is like a collaborator, they don’t do the whole workflow completely automatically, know, like the user is doing some things and the AI is basically as a collaborator adding certain decisions, and so building your Solution on these collaborative, or at least very fine grain version, data models is really key for this to work.

Good, this was everything we learned about user experience, then automation approaches, and this I could go into much much more detail, but I really picked out three things that that I think that are are quite key of how you can use learned models to to really help in these cam decisions. So part segmentation, it’s quite obvious, if you just get Geometry for a computer, this is just like a whole bunch of triangles, okay, what does this now mean. As a person you can look at, so what you see, what you see on the screen is the frame of a removable partial denture, so it’s something that clicks on your teeth, some of these, like what what’s colored in red is like our functional interfaces, they will actually interface with the the the the D themselves, and the rest is nonfunctional. This is something that’s not annotated on the geometry itself, but we can actually quite easily learn this segmentation, based on, there’s a whole slew of different models now, like Point net or graph based models, Transformers, to detect features, and then you can start mapping, typically templates, this could be supporting templates, like process parameters, Machining templates, to these features that you’ve detected.

15:29 We also do this for for Milling, like for example when you have a like these are crowns that are milled out of a ciconia disc, like these crowns need to be, I don’t have a pointer, like these crowns are fixtured with small pins, these pins are added automatically, but there’s proximal surfaces, so the crown actually touches the the teeth on the side, you absolutely don’t want pins on these proximal surfaces, because that then requires a lot more polishing afterwards to get it removed. Like these proximal surfaces is something a human can look at it and immediately knows where it is, for a model to learn that was like non-trivial.

16:08 Good, a second very interesting approach towards automation is similarity, where you can take some geometry, this could be a a whole part, could be like a feature, like a weld line, and then you have a large database of things you’ve done in the past, these are the Million Parts I’ve done before. If you can then somehow find similar Parts in this database, using this, you you typically start with the geometry and then learn some embedding, some like like 20-dimensional vector space, and then you use like elastic search, like vector searching, to find the the 10 most similar parts, and then from those 10 similar Parts you can take the decisions that were made before, supporting decisions, orientation decisions, process parameter settings, and actually retargets them to the new geometry, so this is something that also is is heavily leveraged. Interestingly here it really allows users to tweak to their specific way of working, like you don’t need to retrain the model, you can just look at okay this is how the user changed a certain decision, it’s saved in the database, and then we can just find that example back whenever we have similar geometry, and targeted, very powerful way to to do that.

17:29 Then lastly, like yeah especially in 3D printing, nesting is really quite important, we all want to try to put as many parts on our builds as we can, like in reality it’s much more complex, for example, like in Dental, you want to know about, okay, is the machine and the operator, is there even an operator available to clean out the machine, there’s no operator available, you might want to have the build, for example, like have multiple layers of parts so that the printer runs longer, so there’s no idle time, what about stock availability. And so there’s many many more things you want to add to the nesting optimization to then generate these optimal nests, like these are very hard search problems, you can use like learned metah heris STS to to optimize stuff like that. But so like nesting, very interesting.

18:23 Lastly, and this has been touched a couple times, anyone that has tried to do deep learning at scale, like data is a real problem. I’m just going to give a bit of a of a bit of a diagram on how we looked at the data problem. So everyone knows, okay, you have like this highly well annotated data set, you train these machine learning models, and you plug them in the product, okay, that’s step one. Step two is, oh, I can actually, whenever people make changes in the product, it can feed back to the data set, this is like in your self-driving car, the Tesla, whenever you take over the automation, it’s actually is a, it’s training, for for Tesla, is very similar in the products I showed, where you can even make it such that the models are aware of their uncertainty, I’m not very certain about this scam suggestion, it can then really explicitly almost ask for feedback from the user, which can then Inc like improve the data quality.

19:20 There’s many ways to get to the data, there’s augmentation techniques, there’s like experts we can have it labeled, you can pre-train on public data sets and then fine tune on on highly valuable data sets, you could take information from the production environment, actually know which builds worked and which didn’t work, and very strong signal to then know if the decisions that were made are valid or not, and we can actually use process simulation, we can simulate certain things, search through the simulation space and then use that to generate relevant data that we can train off. Very quickly becomes very complicated, but like this data set, really like a high value well curated data set is the key to make all these all these things work in the long run.

20:14 Good, but to conclude, we were able to successfully roll it out in these two, in these two, in these two Solutions. We really see that we’re able to make machine programming available to many different personas that historically wouldn’t be doing cam programming, we’re able to get the economics of generating machine programs really down, where it’s like viable to actually have like batch size one, where you can create individual Parts using machines, and this same approach I presented here could be applied to like Machining, there’s several companies now active in the space, sheet metal, EDM, mol and die. Good, and of course all these companies support 3MF, both reading and writing, for. Thank you very much.