Transcript: An Exciting Vision of American Manufacturing with Robotics and AI

[00:00:00] Edward Mehr: these are like a 70 year old plane. So they need to have the dies for it, the tooling for it, and the tooling also sometimes doesn't exist.

So for example, for a landing gear door of a certain aircraft, we're looking at four year lead time and millions of dollars before you can get your part. so the aircraft needs to be down on the ground for four years. it's a huge, hugely affect the fleet readiness, for our military. so we're working with them turning some of those four yearly times into days.

​

[00:00:29] Episode intro

[00:00:29] Audrow Nash: There's a big interest in America to reshore manufacturing. Now, to do this, do we just bring the jobs back to the U. S. and overlook that the reasons that those jobs may have moved may still exist? Or do we invent new ways of doing things and leverage our strengths in technology and innovation.

This interview is with Edward Mehr, and he is firmly in the second camp. He wants to invent new ways to do manufacturing, and he's well positioned to help in this effort. He's the CEO and a co founder of Machina Labs, and they're reinventing metal bending with robotics and AI. In some cases, taking something that literally is taking us years to make, and they're doing it in just days.

It was an awesome conversation and got me excited about the potential of robots and AI to greatly improve how we do manufacturing.

You'll like this interview if you're curious about robotics and AI in manufacturing, interested in how great robotics companies are built, and if you're interested in AI for modeling complex phenomena like metal bending.

Without further ado, here's my conversation with Edward.

​

[00:01:51] Introducing Edward and Machina Labs

[00:01:51] Audrow Nash: Would you introduce yourself?

[00:01:53] Edward Mehr: Yes, my name is Edward Mehr. I am CEO and co founder at Machina Labs.

[00:01:58] Audrow Nash: Hell yeah. Would you tell me about Machina Labs?

[00:02:01] Edward Mehr: Yeah, so we are, working on the next generation of manufacturing floors. the real problem we're trying to solve is that today, if you want to build a physical part, you pretty much have to build a factory that is very specifically built for that part. A lot of tooling, a lot of machinery that goes into factories, are very specifically designed for the geometry, for the material you're trying to manufacture.

And that severely limits in terms of the amount of CapEx investment you have to make to build a part,

[00:02:36] Audrow Nash: So it's far more expensive because you have to invest in infrastructure to build it. Okay. Or equipment.

[00:02:42] Edward Mehr: What we're trying to do is, can we develop technologies that allow you to, make a part A today out of design A and material B, and then switch to design C and material D tomorrow without having to change your factory.

[00:02:56] Audrow Nash: I love it. Hell yeah. So how, are you approaching that? Because that, sounds very interesting to be able to switch so quickly. And I've heard that we're going from like high volume low mix meaning a lot of the same thing being made to more of lower volume with high mix and this sounds like it's working towards that what are some like what do these manufacturing techniques look like

[00:03:26] Edward Mehr: Yeah, it's funny you mentioned it going from high volume low mix to high mix low volume. I think actually the ideal combination is high mix, high volume, right? because always people think about it in terms of it's a compromise, right? Because that was

[00:03:42] Audrow Nash: a trade off

[00:03:43] Edward Mehr: was given as a trade off, right?

Like you need to have high volume and low mix or low mix and high volume. What is it? But If we want to think about it fundamentally, what is the technology that needs to be there to enable both high and low volume and high and low mix? that's actually what we're thinking about. but to answer your question directly, if you look at history of manufacturing, and if you go Maybe 200, 300 years ago, where manufacturing first started as a craft, not to start it, but like it was mostly done as a craft.

We actually had a lot of flexibility. We had these people, we call them craftsmen or blacksmiths. and they had a very creative mind. They had very limited set of tools like hammers and chisels, maybe a set of tool of tools. 10 different kind, but then you could one day go to them and say, Hey, I have this piece of rod, can you turn it into a sword?

And then they would hammer it into a sword, right? They will figure out with their mind, how they're going to use their limited tool set they have to make a, sword. And the next day you can go to them and be like, I have this piece of sheet of metal, and I want to make a shield. And then they would use the exact same tools to use the sword, but apply it differently in a creative way to make you a shield.

So they were actually very flexible. The challenge was that a craftsman was a very much of a learned skill. You had to do mentorship. You had to, to go learn it from somebody else who has done it. so it would take a while before somebody would become expert in it. And then more importantly, humans, have limitation in terms of throughput.

So you could make you maybe make one sword, two sword, three swords a day. but beyond that, You couldn't, right? So with industrial revolution, there's a change happened where we were like, okay, craftsmen are great, flexible, but they can't have a throughput. So we build these machines that can make the same thing over and over again, right?

Because we were not creative enough or didn't have enough technology to develop the same thing that the craftsman does. We could make very constraint machines, hardware constraint machines that can do the same thing over and over again. but we couldn't make flexibility because we couldn't Replicate intelligence that the craftsman has.

So for the past two centuries or more, up to even today, manufacturing morphed into make the same thing over and over again. Even today, if you're like, for example, an automotive business, a car business, the main way you can make money is create one car that everybody loves. And then make a lot of it.

That's how you can make margin. because manufacturing is very much dependent on the type of parts you're making and you cannot easily change it. Look at Tesla, they're planning to make 5, 6 million of Model Y a year sometimes. sometimes. and that means that, the game is still the same as when industrial revolution started, right?

You have to make the same thing over and over again, to make it, profitable. Now, today. Advancement in AI and Robotics. We could replicate what a craftsman does at a scale. You have robotics that has the same kinematic freedom as a craftsman, right? Maybe even more dexterity, maybe more precision, higher force that they can apply.

But then with AI, we also can replicate what happens in the mind of the craftsman. We can figure out how to use different tools creatively to replicate different types of processes. So I think now we're, the two ingredients To make a flexible manufacturing system exist. So to answer your question directly, our systems use robotics and artificial intelligence to almost replicate what a craftsman does, but in a scalable way.

Because once you build one robotic system can do a craftsman, I can just replicate it and have thousands and thousands of those systems that does the same thing. Now you get flexibility, but you also get scale through, through scaling these robotic systems.

[00:07:33] AI in manufacturing

[00:07:33] Audrow Nash: Okay. So I'm interested in what you are referring to with AI for this. what are you, what kinds of things can we use AI to come in and help with manufacturing and maybe some simple reasoning or about what to do next or what to do? how are you using ai?

[00:07:57] Edward Mehr: AI, the definition of AI has changed over time

[00:08:00] Audrow Nash: Oh, definitely.

[00:08:01] Edward Mehr: Since its inception. If you look at 1960s, and I come from a computer science background, if you look at 1960s, intelligent AI systems were actually rule based systems. We'll call rule based systems intelligent AI systems, like regular programming.

And then over time, as we develop new techniques, new empirical techniques, where we can find patterns in the data and, use those patterns to actually do an object, do a task that a human does more easily and more automated. Then it starts shaping the definition of AI to more things that the humans can do.

there are a whole array of tasks in manufacturing that falls within different categories. Up to today, most of the automation has been what I call classic, classical AI, where you define the rules and the system just follows the rules. if you look at the type of manufacturing we talked about, where a craftsman can figure out a sequence of events.

can look at a piece of metal and as it's hammering it, figure out what is the next steps that it needs to do. That's the type of AI we're talking about. It's not following a set of rules, it's finding patterns in making a task done. And that task might be forming a sheet of metal into precision from a flat sheet.

What are the set of, process parameters? What are the set of, sequence of tasks that needs to be done to get a flat sheet of metal all the way to the final product? You cannot easily define it. You cannot define it. It's not a rule base. That's why it has been a craft. You have to learn it from somebody, mentor.

but with AI, we can actually capture what is happening in the process using different ways of capturing data. And then it starts identifying what are the patterns that the craftsman uses to get to the right part. And that's what we're talking about here. it's similar to some of the things you might see in, in the, in.

Discussions that you see around AI outside of manufacturing, how can a ChatGPT then do similar to human reasoning. That's the type of AI we're talking about here. That being said, we also use a lot of more earlier versions of AI, where, for example, as I'm forming a part, can I look at the Data from the sensors and find an anomalous behavior where it might lead into a defect into the part.

So classical, anomaly detection using empirical data. Those are the things that I also use. But the main enabler is, figuring out how a craftsman forms a part or crafts a part. What are the set of rules and, guidelines that we need to develop so that the robotic system can actually have the same performance as a craftsman?

[00:10:48] Audrow Nash: Okay. I would love if you can take me through a concrete example, because I, feel like I understand a bit from how you're describing it, but I want to see like in a metal forming or something like this, how do you, how are you using this to achieve a desired part?

[00:11:05] Edward Mehr: Yes. So for folks who might have seen our videos, you will see two robots, for example, forming a sheet of metal from flat sheet. So you have a flat sheet, two giant robots on two sides, and almost like a potter, they're like two stylists looking endeffectors on two robots on two sided sheet. They come in and pinch and deform a sheet of metal slowly into shape.

Now, when you look at it, it almost looks like a very much like a CNC process, right? Where you're almost like doing waterlines of the part and deforming the part in waterlines until you get the final part. But in reality, it's not just a CNC process. It's a lot of different changes in the process parameters, but also the path the robot takes is sometimes you might go in the center of the part and start forming the center of the part, and then go back to the outside and form the outside of the part.

So why do we have to do that? Why can't just a simple heuristic CNC path doesn't get you the right part? It's because, as you're forming the part, the sheet moves in unpredictable ways. And also, there's a process, there's a phenomenon called a springback, where you might form a sheet to a certain length or certain depth.

After the robot moves away, the sheet jumps back into a different shape, right? you need to account for all these movements, all these springback phenomenons that we talked about to form the right part. So you end up, the path that the robot ends up taking to form a part that's like less than a millimeter accurate to the design is a very non intuitive path.

You can, a human cannot just think of it. and

[00:12:39] Audrow Nash: you can't think of it all at once, but you can probably iterate to get there.

[00:12:43] Edward Mehr: you could iterate to get there. Yes. And that's what, we actually generate the data. So for right now, what happens is, you run maybe a heuristic path that you think, oh, this might work well. And then you scan it and say, okay, This area was five inches off, that area was two inches off, compared to the CAD.

So I can make a slight adjustments to get it a little slightly closer, and then run another one, and then see the results, and run another one. But if I could have a model that would tell me, in order to get to this part, this is the non intuitive path you can take. And you didn't have to do those trials. you can significantly improve the performance of the system. So what we're doing at Machina is like we're starting with human intelligence and creating a very seamless interface that humans can iterate and build these parts while we capture this data and use this data to build models that slowly improve the efficiency of the humans.

in early days, we would do 25 trials to get a part done. Now we are down to five, six trials. And the goal is get to the point that right off the bat, you get

[00:13:45] Audrow Nash: Goes to the right thing.

[00:13:46] Edward Mehr: done. Yeah.

[00:13:48] Audrow Nash: That's really cool. Hell yeah. So the way I'm understanding it is you, so in this metal forming task where you have the two robots on the one on each side of the sheet and they pinch so that they can push in to deform the metal in some sort of way to make some complex part, you have a part that you want to make and you are iterating in some sense so you like you'll push you'll pinch in some way and deform it and you do that a few times and you say this is what I expected this is what it actually looks like when you scan it or something to understand the state of the part and then you can keep steps until you converge on what you actually want the part to look like then you're also generating data to make it so you have a better, you're generating data so that you can use that to have a better model of how all of the things you're doing to the metal sheet or whatever the sheet is, you're understanding what happens when you apply an action to that sheet so that you could be more efficient with your actions,

Then you're more likely to be reliable and then you're also faster because you are better able to make the part in fewer actions and this kind of thing is that?

[00:15:19] Edward Mehr: No, you're exactly,

[00:15:19] Using data for better models than physics-based models

[00:15:19] Edward Mehr: you're basically building models that gives you an understanding of what is the physics that is happening underneath. And use the data to do that.

[00:15:28] Audrow Nash: Rather than directly modeling it because it's very complex because of all the really complex contact forces

[00:15:35] Edward Mehr: Contact forces, friction, so if you want to, traditionally people when they would do this, They would use physics based modeling, right? people think of finite element analysis or computational fluid dynamics. Those are methods that allows you to simulate a physical phenomenon. The challenge is that with a lot of these processes, first of all, FEA is very slow,

[00:15:57] Audrow Nash: What was that?

[00:15:58] Edward Mehr: FEA or Finite Element Analysis, which is the physics based way of doing these things.

[00:16:02] Audrow Nash: Yeah, you simulate like a bunch of points and you see what happens and you do, it's a differential equation to solve

[00:16:08] Edward Mehr: You apply physical laws to, to figure out, okay, what is happening. One challenge with that approach is that it is a slow. So it requires a lot of computation to, to do it. We, in early days of our company, we would form these parts that would like to take 20 minutes to form. And if you wanted to simulate it using physics based models, it would take us, on a 27 core machine, it would take us a week. So we're like, okay,

[00:16:32] Audrow Nash: Not feasible. Yeah.

[00:16:33] Edward Mehr: Not feasible. I would rather just run the part,

[00:16:36] Audrow Nash: Yeah.

[00:16:36] Edward Mehr: Let's see the results in real life, as opposed to simulate. Simulate it with nature.

[00:16:40] Audrow Nash: like heat people's homes with, computers that are running at full, bore the whole time.

[00:16:45] Edward Mehr: So we, made it, we changed the parts. We're like, okay, maybe can we do this, build these models empirically? And the other challenge with physics based models is that even if they're fast, There are not accurate.

To your point, you might not be modeling everything that's in there. For example, there might be some miscalibration in a robot that you're not modeling. There might be some, friction forces that causes some issues. There might be some adhesion between the end effector and the material that you're not modeling, or you don't, you're

[00:17:11] Audrow Nash: The model just isn't good enough. To do the task. like one thing that really struck me, in my electrical engineering education is so voltage equals resistance times current, but so V equals IR, but not at high frequencies. and then so the model breaks down. So the model is correct up to some, in some environment, or at least pretty close to correct in some environment.

And so if you're doing complex things on a part, I would imagine the physics based models just don't capture all the phenomenon very well.

[00:17:45] Edward Mehr: Yeah, you're exactly right. And that's once there's a need to build these empirical models, they're faster to inference, right? So you can get the result of it really fast once you build it. And then because they're based on physical data, you're almost taking into account everything that's happening in real world.

The key though is, as you're building these robotic systems, you need to build a system that allows you to expose that data you need to build these models. And that's what some of the architectural decisions we had to make is that, how can I capture the data that I need at every millisecond of this process?

So I can model it. and affects your product development and what kind of robotics you want to use and what kind of control loops you want to use.

[00:18:22] Audrow Nash: Okay. So I just, that is very interesting. I want to get more into that, but just so I understand at a high level, as I'm understanding things now, you have two parts, really. one is the iterative working towards a task using your existing model. And the other one is the, I've captured data. Let me learn a better model part. So you have those two things. Okay. so

[00:18:51] Capturing data + modeling

[00:18:51] Audrow Nash: how did you design it for, how did you design your system so that you can make an efficient transfer to your data part of this so that you can learn a better model? Because that's very interesting.

[00:19:02] Edward Mehr: Yeah, so there is, the choice of your architecture, your hardware architecture greatly affects your ability to, get data. So you can imagine, for example, in a stamping operation, which is a traditional way of forming sheet metal parts, where you have a giant press, and you have dyes that are male and a female dye, and you put a sheet in between and you stamp it.

Not a whole lot of opportunities to capture data, okay, we're going to

[00:19:28] Audrow Nash: it's like start and end

[00:19:29] Edward Mehr: It's a start and then you can capture that Some kind of an input to the to the either it's a hydraulic press or a servo press the current or hydraulic pressure Or whatever throughout the time, but you don't really you cannot really know what is happening to your part in a very granular way Because it's inside a very destructive stamping press so with our process because we're incrementally forming it we can capture A lot of little details, like what is the forces, every once in a while, scan it, what is the response, on both robots, how much current we're putting through the motors, how much deflection we're looking at each of the robots,

[00:20:08] Audrow Nash: wow, so you're doing this on the full system. It's not just Okay, that's, really interesting. I was assuming you were just, I, what I had assumed is that you just have the sheet and then you have a model that says we applied this to this. force this point force at these locations and then we scanned it and this was the result and that would inform the model but going down to like current on motors on the arm is really awesome because i guess they're non linear in their response most likely and so then modeling that actually probably makes a lot of sense Sounds hard though.

[00:20:43] Edward Mehr: yeah, so the robot itself, and you're in robotics, so you fully appreciate this, is depending on the pose, you might get different, deflection profiles, right? If a robot's completely, extended out. It might deflect much easier with less force. So the K constant for it is much, much, more flexible than like when it is in, in a very stiff position.

So we need to actually can take into account both the system itself, how it's responding to the process, and then also the material, deforming, what forces you're required to deform the sheet. But what kind of information you're going to get, what kind of springback you want to get. So you want to capture that at every step of the process.

And the more data, the more granular you can capture the data, the more easier it is for you to make the model. But more importantly, the more data you have to build a model, because a lot of these processes require a lot of data. So if you get had to start an end model, you have to make a lot of parts.

But if I could get deflection at every step. point of couple hours that

[00:21:46] Audrow Nash: You get a lot more data. Yeah

[00:21:48] Edward Mehr: I have a lot more data.

[00:21:49] Audrow Nash: It's a lot easier to converge to a good model with that I would imagine.

[00:21:55] Edward Mehr: Yes.

[00:21:55] Audrow Nash: One question just for my understanding because I don't know too much about metal forming You are, when you're metal forming you constrain all the edges of the sheet of Metal that you're going to be forming, right?

[00:22:11] Edward Mehr: Yes, you could. we do

[00:22:11] Audrow Nash: it's sitting in a picture frame almost, right?

[00:22:14] Edward Mehr: Yeah, you could do that. that's what we do today. That's what, that's another parameter. You could, modulate the amount of basic pressure you're putting on the boundary.

[00:22:22] Audrow Nash: So then when you do that, you're pushing into it and you're pinching part of the metal out. Are you, so what I assume is that the metal gets a bit thinner where it's stretched out. and so I would imagine that's complex. I don't know between, so in a finished part, how significant is the thinness, the thickness of the material?

Changed when you push it out, quite a bit. Does that, because that would be another dimension in the modeling, I would imagine the thickness of the metal part now

[00:22:59] Edward Mehr: Yes, so it depends on what path you take to deform the material. So if I start from a flat sheet and just, create a 60 degree wall angle right on my first layer as I'm forming the part, then the calculation is relatively easy. It's basically a volume, yes, it's a volume preservation, basically, law, right?

Okay. I. Got, start from a X millimeter thickness sheet. I deform it to 60 degree wall angle. So it's going to reduce by cosine of 60 degrees times the original thickness right? So simple, trigonometry will give you the answer. But if I, but, if, for example, if I did first a 45 degree wall angle and then push the 45 degree into 60 degree wall angle,

[00:23:47] Audrow Nash: Now it

[00:23:47] Edward Mehr: my thickness profile is completely different, right?

It's cosine of 45 times cosine of 60, which is a smaller number. So actually you get less. Less thinning, right? So another, so if you add thickness preservation or set thickness targets for your geometry, your final geometry, which you can, that also affects in what order you want to form the part, right?

You might want to form intermediate geometries and then push those intermediate geometries to the final geometry.

[00:24:19] Audrow Nash: Gotcha. Very interesting. Okay. going back to the data that you're generating. if you are observing kind of the robot state, also, robot, I just, I'm, imagining your state space is blowing up and this becomes a very complex equation because you use two, seven degree of freedom robotic arms.

Is it right?

[00:24:46] Edward Mehr: Correct.

[00:24:47] Audrow Nash: Okay. And you're observing the, current at each joint in the seven degree of freedom robot arms.

[00:24:56] Edward Mehr: Yeah, you could, you have access to the current at each edge. Yes.

[00:25:01] Audrow Nash: I'm imagining it's a very challenging optimization because it's very high dimension. And maybe you use something like neural networks that are very expressive to fit this. but then you might get a big risk of overfitting and it might be tough to generalize.

just, tell me about some of those challenges of working in a high dimensional space and how you guys deal with this.

[00:25:25] Edward Mehr: Yeah, so when you're in high dimensional space, you have two challenges. basically you have two routes of making models that are accurate. One is generate a lot of data,

[00:25:34] Audrow Nash: Which you're doing.

[00:25:36] Edward Mehr: which, is something you can do and we can, we are doing as well. Or in the short term. in the earlier days where you don't have enough data, you can also try to break down the problem, basically do physics informed modeling.

[00:25:48] Audrow Nash: Yeah,

[00:25:50] Edward Mehr: So you can start

[00:25:50] Audrow Nash: you reduce your space for

[00:25:52] Edward Mehr: introduce your space.

[00:25:54] Audrow Nash: by leveraging things like physics, simulation. Okay, I see.

[00:25:58] Edward Mehr: For example, you can say, okay, if my robot is deflecting, I can either create a deflection model based on the currents and the joint angles, right? and that would be my space. Or I can just say most of the deflection happens in each joint and, then model each joint differently and then calculate the total deflection based on the configuration of the joints.

So now you're taking advantage of kinematics and physics to say, okay, I can simplify. I almost feature engineering, and they, call it in the data world, like I am informing it. Maybe there's a better way I can solve half of it for you as long as you tell me how each joint deflect.

[00:26:41] Audrow Nash: Yeah,

[00:26:42] Edward Mehr: and then, so those are some of the things we did early on.

[00:26:44] Audrow Nash: So you're compressing your dimensions, basically, so you have fewer dimensions, and it's more tractable with whatever data you have. How do you go from zero data to some data? Do you start with physics based and you just poke it a few times and see what happens and then start to bootstrap a model or how is that?

[00:27:03] Edward Mehr: That's exactly what we did, right? Do like back of the envelope, not even physics space, back of the envelope calculation initially. I remember our initial deflection compensation model wasn't even in joint space. It was in Cartesian space, which has nothing to do with the robot, but we're like, okay, maybe

[00:27:20] Audrow Nash: so I would have approached it too.

[00:27:22] Edward Mehr: like they're like different in z direction as opposed to x y, so I can do a Cartesian based, deflection model. That took us a long way until we start seeing tears in very complex parts, and then we're like, okay, This is model, this is pose dependent, so we need to start, increase the complexity of the model.

But at that point, you have enough more data, enough data to start looking into what that model will look like. Yeah, the key,

[00:27:48] Audrow Nash: Oh, go ahead.

[00:27:50] Edward Mehr: I think the key for basically, I was listening to this podcast with Andrej Karpathy about why, or no, actually it had to be with Ilya, about why, ChatGPT exists, but not, A lot, it's much harder to do foundational models for robots. key is because you need to gather capture data. And in order to capture data, you need to operate a huge fleet of robots. Now, yes. And then this, means that you need to figure out a way to find an application where you can create value based on heuristic simple models to begin with, but that application after the data can have a potential to significantly improve.

but you need to find that application. That's a tough part. Find an application where even today with heuristic models, you can provide benefits.

[00:28:39] Audrow Nash: I think Electric Sheep, who was the first interview on this podcast, is doing a great job of this. They are doing lawn mowing. With very small mowers. And I think that's a very strong application for this low risk, simple, just like they're doing localization and path planning and things. And they're learning a lot of that.

but I think there are probably a lot of other spaces that could do a similar approach. So what I would suspect is that realizing, okay, data, is really important. The more data you have, the more expressive your model can be and the better it will perform.

so I imagine. That realizing that you decided to grab all the data you could. and then you've slowly been increasing the number of dimensions that you're actually using in your optimization. Is that true? And then also, how, where are we now in terms of, like, how much is in the models? How expressive are the models? How many dimensions are you including in the models? This kind of thing.

[00:29:45] Edward Mehr: So yes, you're right. So early days, basically build a lot of stuff based on heuristic or back of the envelope calculation or

[00:29:53] Audrow Nash: Heuristics still.

[00:29:55] Edward Mehr: Yeah. And, and then we start forming parts, like I said, early days, we had 25 trials before we get a part, right. Now we're down to five or six.

And initial models we used a lot of modeling, the techniques we use were so very simple. Like you could even start from regression. because once you break it down into a simple thing, you can think of, let's say, a joint deflection as a simple, one, one degree polynomial, right? because it's mostly

[00:30:26] Audrow Nash: simple models for these things.

[00:30:28] Edward Mehr: right?

So, we start from there. and I think last year, so to give you a little bit of concept, like last year, it's actually a paper that we, published with Northwestern around predicting the forces in our process using graph based neural networks. so that's like the, state of the art last year.

and it, starts allowing you to do cool things. for example, in that, paper, we use transfer learning.

[00:30:59] Audrow Nash: what? Transfer learning?

[00:31:00] Edward Mehr: Transfer learning, and the way it works is that instead of trying to make a lot of assumptions about our system, as you can imagine, you mentioned yourself, is that there are a lot of parameters.

What if your calibration is even wrong to begin with, from system to system? How does that affect the whole

[00:31:14] Audrow Nash: Or you could lump them into additional dimensions that you're doing,

[00:31:18] Edward Mehr: Exactly, you could lump them in an additional dimension, but then you need more data. So what we ended up doing to answer your question, in the short term, what can you do to actually improve the performance of the models?

We started using transfer learning. We said, okay, with every system, first form 10 layers, and then retrain the model that you have on just that 10 layers in that machine with that process parameters. And then predict what's going to happen for the rest of the part and optimize for it based on that.

And then we got much better results because now you're learning the extrinsic very quickly. different, specific features that each machine might have or each configuration might have and learn those in the first few layers, apply it using transfer learning, and then, you've got to be more accurate.

So that was the state of the art last year. Now this year, we are going back again at that bigger prize. Can I go all the way from this is input geometry, what is the final design if I go through a certain set of parameters? So that's what we're working on. I think we have enough data now to capture that.

so, stay tuned. I'll let you know how that model goes. And we're pretty open. We publish most of the stuff we, we do. Like I said, that paper is going to be in NAMRAC and it's going to be presented.

[00:32:30] Audrow Nash: Hell yeah.

[00:32:31] Edward Mehr: But that's, the, state of the art in terms of, the modeling today.

We're still using neural networks, but we're trying to basically go with the bigger scope of the, model and not do small feature engineering or, do things that are like physics informed.

[00:32:48] Audrow Nash: Very cool. Yeah. You're removing a lot of the heuristics and seeing if you can learn them and if that will be better. And then transfer learning is very clever, I think, because you can learn a good bit and you can make it general, but then you can learn the very fine details on each specific machine, or with a new material, or that kind of thing.

[00:33:11] Edward Mehr: It's funny, craftsmen do that too. So I used to do sheet shaping myself, and we used to, I used to work, at the shop, in Pomoda, we would do, panels for hot rods. And you do it under a, PowerHammer, you form it. But a lot of times you get a material, you don't know what it is.

they say, okay, this is a mild steel, maybe. And then you hammer it a little bit and you get a feel of it. You were like, okay, this is how this material operates. And then you, estimate what you need to do and how long it's going to take you. and I think it's the same thing with transfer learning, get the robot to form a slight little part or like few layers of the first part, and then capture that and use that data to figure out and form the rest of the part.

And that's something that humans do. I had a, blacksmithing teacher, you learn all these little tricks where oh, if you go to a junkyard and you want to buy good metal, we used to carry these hand pocket grinders.

[00:34:04] Audrow Nash: Oh, wow.

[00:34:06] Edward Mehr: and then you grind it and based on the amount of sparks it makes, you can know how much carbon it has.

[00:34:11] Audrow Nash: Oh,

[00:34:12] Edward Mehr: and the more carbon it has, it's usually, yeah, a stronger material. But this is, these are all concepts, like you try the material and you see the feedback and response of the material, and you know exactly what you're dealing with. So this is what craftsmen used to do, like mental patterns that created for a long time.

I think now with robotic and AI systems, we can actually replicate that, and maybe it's a faster way to scale these processes.

[00:34:33] Audrow Nash: Yeah, it's super cool. Hell yeah. Tell me a little bit. So you mentioned you're publishing research papers, which is very cool. I'm glad you're contributing to the collective knowledge of the whole robotics community.

[00:34:45] About Machina Labs

[00:34:45] Audrow Nash: tell me a bit about your team and Where are you?

I know you're in LA, but tell me about, just about your company in general. how many people you are? How large is your space? What kind of things are you working on? is it still very early researchy days or are you having customers and all these details?

[00:35:09] Edward Mehr: Yeah, for sure. so we started in 2019. and, we started in a facility. It's also a, defense contract manufacturer. That's our landlord. We got the back portion of this facility with actually a little bit of shares. We paid the landlord rent in shares and they're very supportive. but since then we can, we have, we can expand it in the same facility.

We have 35, 000 square feet in this facility. we've deployed 11 robotic cells downstairs, so 22 robots in total. and then we just acquired another space. It's a 66, 000 square feet. It's two miles away. So now we have around 100, 110, 000 square feet in total. we are slowly moving into data space as well.

but the goal of that space is not so much house our robotic cells to increase the, rate of manufacturing.

[00:36:01] Audrow Nash: More data and everything else.

[00:36:02] Edward Mehr: yes. So we are, hoping to get to a point where, we can manufacture a manufacturing cell in a month. so, we can get 12 manufacturing cells a year,

[00:36:12] Audrow Nash: It's so meta. That's great.

[00:36:14] Edward Mehr: Yeah. So, that's, the company's right now close to 70 people. We're planning to get to around 130, 140 people by the end of next year.

[00:36:23] Audrow Nash: Wow. Damn.

[00:36:24] Edward Mehr: so we're growing rapidly. the cost, the product is

[00:36:28] Audrow Nash: by the end of next year? So in the next, what is it? I don't know, 18 months or something?

[00:36:34] Edward Mehr: year and something. Yes. mid next year, I think that's, Yes.

and yeah, so the company, almost has been doubling the staff every year. and, it's already in customer's hands. early days of, our company was about making parts. one thing I learned from my previous job was that do not lock yourself in the lab. And by the time you're done building it, the product is already obsolete.

so we ended up even the first time, and even we didn't have robots, we were lucky enough to get customers to be like, okay, what are some of the complex challenges you have on sheet metal? Send it to us. We're going to, we're going to work on those. so I think NASA and Air Force were one of our early customers who provided us with the challenges they had.

so since day one, we had Customer Paying Parts, even our first year of operation, I think we made 300, 000 with me and my co founder, so it was, a good amount of interaction very early on, and we've been able to like almost double the revenue or more than double the revenue every year since then,

[00:37:42] Audrow Nash: Hell yeah.

[00:37:44] Edward Mehr: the system right now, we do manufacture parts in our facility, but that's usually in the goal of eventually them deploying the system in their, in their,

[00:37:54] Audrow Nash: Do you want to sell work cells effectively? Oh,

[00:37:57] Edward Mehr: Yeah, we already have deployed ourselves into the customer facility. So for example, one of them is at Warner Robins Air Force Base, working on building, components for, sustainment and repair of aircrafts. they have an interesting challenge in the military is, they have, hundreds of weapon systems and defense systems that are, some of them, 70, 80 years old.

think about B 52. and these are sheet metal airplanes, right? And every time a component, let's say a landing gear door gets damaged, they either have an inventory of it, but in a lot of cases they don't, because these are like a 70 year old plane. So they need to have the dies for it, the tooling for it, and the tooling also sometimes doesn't exist.

So for example, for a landing gear door of a certain aircraft, we're looking at four year lead time and millions of dollars before you can get your part. so the aircraft needs to be down on the ground for four years. it's a huge, hugely affect the fleet readiness, for our military. so we're working with them turning some of those four yearly times into days.

[00:39:05] Audrow Nash: Hell yeah.

[00:39:06] Edward Mehr: Yeah, so that's like one of our biggest customers. it's deployed there and we're working with some of the other folks in automotive and aerospace as well.

[00:39:15] Machina Lab's funding

[00:39:15] Audrow Nash: Okay. 70 people, two places in LA, 22 robots. How are you guys funded? So you mentioned you're already getting revenue. but I assume you've taken some investment rounds or how, have you guys been funded up to this point other than the revenue you're making?

[00:39:34] Edward Mehr: Yeah, so we're a venture backed company, which has its own challenges and good, things and bad things. to this date, we've raised about 45 million in venture funding, right? So we're, very,

[00:39:49] Audrow Nash: What round, have you

[00:39:50] Edward Mehr: We just raised our Series B last year

[00:39:53] Audrow Nash: Oh, congrats. Oh yeah,

[00:39:54] Edward Mehr: last year. Yeah.

[00:39:56] Audrow Nash: That's fantastic because I hear that it gets much harder, like B and C are the two hardest funding rounds for robotics, I think for hardware in general

[00:40:05] Edward Mehr: yes.

I think that especially last year was pretty tough because, once you get to B and C, you need to have commercial traction. you need to show you have a good revenue and you also have good margins that is suitable for a venture funded business. That's the one challenge with venture funded business is that what is their alternative?

The alternative is a software or ChatGPT that it's I don't know, what is the margins there? Like 80, 90%, whatever it is today,

[00:40:33] Audrow Nash: I don't know.

[00:40:33] Edward Mehr: For us, so we need to be able to compete with that. and it's much tougher for robotic companies with CapEx expenditure to get there. Out. But I think we have been making a lot of interesting choices and hopefully wise choices that allows us to our margin to stay up so we can compete with those type of, those types of businesses, but I guess tougher for, as you pointed out in B and C because economics is the main driver of raising a C round or a B round.

[00:41:00] Audrow Nash: there are a lot of good, I don't know what the word is, like headwinds or forces or something. like we're reshoring a lot of manufacturing is what it seems. and so I would think our government is very supportive of what you guys are doing, which I think is wonderful. And then, I don't know there's just probably a big interest like just seeing what feels like the zeitgeist on X it's like manufacturing is becoming cool wasn't the case years ago and I think it's a wonderful thing for our country in general.

[00:41:36] Edward Mehr: I agree. I agree. I think this year, especially, I think the, leaf is turning. I think last year was still a little bit tougher. Everybody was in holding mode. I think now there's a lot of excitement about manufacturing, but I think also at the same time that puts the,

[00:41:52] Audrow Nash: High expectations

[00:41:53] Edward Mehr: Expectations on it, which is in a good way under robotics and hardware and manufacturing companies to really go about this in a smart way.

Yeah.

[00:42:02] Audrow Nash: Especially with cash being so expensive relatively now because then yeah you have to be pragmatic or you go out of business this kind of thing

[00:42:11] Edward Mehr: Yes. So turn this excitement and make it sustainable, right? So, that all these companies that come in, obviously a lot of them will be successful, some won't. but I think it's new grounds. we have to think about our business models are not going to be the same as SaaS business models.

Our, the way we're going to make margins is going to be different, but we need to pave those paths. Like it's interesting that even investors. Don't know exactly what the model will look like, but if the margins are not 60 70 percent at the end of the day, it's tougher to compete with some of the alternatives they have, so we need to be very, we need to think about it extra hard and be very adaptable and agile here.

[00:42:50] Audrow Nash: Yeah, so what was your valuation?

[00:42:53] Edward Mehr: Yeah,

[00:42:54] Audrow Nash: I don't know if that's public, but

[00:42:56] Edward Mehr: yeah, It's not public, but we raised relatively with good,

[00:43:03] Audrow Nash: Didn't trade that much equity for everything.

[00:43:05] Edward Mehr: Yeah, so we, have been every round, we at least doubled or tripled our, more than double and sometimes tripled our valuation, from the beginning. I won't be able to give you the exact amount,

[00:43:18] Audrow Nash: Yeah, ballpark would be

[00:43:19] Edward Mehr: it has been increased.

Again, we

[00:43:20] Audrow Nash: has been increased. It was a good round, basically. Good, yeah, because a lot of companies have been seeing down rounds, I think, which just, I don't know, cash is becoming tighter, or less. I just, had an interview,

With a person in venture capital and they were saying that in 2021-2022 kind of thing, it was just like everyone was throwing money at everything.

And so people raised huge up rounds and then, now the expectations are coming more back down to earth. And then, That has created some friction for a lot of companies and a lot of investors.

[00:44:06] Edward Mehr: Yeah, no, you're absolutely right. I think, we were lucky that every round is more than double the

[00:44:12] Audrow Nash: Yep. So you guys are doing great. Hell yeah.

[00:44:15] Edward Mehr: I think part of that is as tempting as it is to get a very high valuation in early days, I think it is very important to, almost pace your progress and set milestones in a way where you don't set yourself up for a down round. I think a lot of folks come in and, tell very amazing stories, which is great. I'll get the biggest one I can, and then it's okay, the milestone now I'm setting up because if the valuation is high, the expectation is high too. So the milestones you set up for yourself for the next round, if you don't meet those, then You know, the valuation drops and that's not a good thing for your employees that are stakeholders in the company.

Obviously not a good thing for your investors as well, but more importantly for the employees who are, leaving jobs at Microsoft and Google to join your company. I want to always show them that, this is a better alternative than staying in those places. If you have a downturn, that's not a good thing.

[00:45:11] Being in LA

[00:45:11] Audrow Nash: How, so another thing that's interesting to me, you guys are in LA. Tell me about the LA, and actually, how, how close to, the major part of LA are you, far out east, or where, where, are you in the LA region?

[00:45:32] Edward Mehr: Yeah, if you're familiar with it, I think we're in the Valley.

[00:45:35] Audrow Nash: Oh, hell yeah. So

[00:45:36] Edward Mehr: it is Los Angeles, the city. and, you're probably half an hour away from downtown, if there's no traffic.

[00:45:44] Audrow Nash: If there's no traffic. Yeah.

[00:45:46] Edward Mehr: Yes. for like maybe 45 minutes away from SpaceX, for those people who are like anchoring a lot of harbor companies around the SpaceX in South Bay.

So we are on the other side of the LA. So South Bay is traditionally have been aerospace manufacturing hub. We are in the other matter of fact, in the other manufacturing hub, Chatsworth, which is North end of the LA area in the Valley. We have companies like a lot of aerospace companies and machine shops here as well.

So it's a big manufacturing hub. it's one of the other, basically it's the other manufacturing hub,

[00:46:18] Audrow Nash: What's the other one from that?

[00:46:21] Edward Mehr: South Bay, like

[00:46:22] Audrow Nash: Oh, other one in LA. I see.

[00:46:23] Edward Mehr: yeah, El Segundo, Torrance, Hawthorne, that region, which is like 45 minutes away from us, yeah.

[00:46:31] Audrow Nash: And so do you think it's a very good thing for you guys to be very close to like relatively close to SpaceX and a lot of military bases? is it a really good spot?

[00:46:43] Edward Mehr: I think so. I think we have a good ability to attract hardware talent here. Software is a little bit tougher, right? Because

[00:46:54] Audrow Nash: Oh, they're in SF or in the

[00:46:56] Edward Mehr: Area, or maybe East Coast a little bit, Boston,

[00:47:01] Audrow Nash: Oh,

[00:47:02] Edward Mehr: and New York. So it's a little bit tougher on the software side. But overall, I think, the hardware talent is much more accessible here.

[00:47:11] Audrow Nash: Gotcha. Very cool. Let's see. So pivoting a little bit,

[00:47:16] Why make a company that is 10x better than alternatives

[00:47:16] Audrow Nash: why haven't we seen many big and successful robotics companies yet? So

[00:47:25] Edward Mehr: Yeah, I think nobody really has figured out what is the right business model for robotics companies. I think for a robotic company to be successful, for any company to be successful, I think, especially a startup, you need to almost provide 10x improvements. the reason for 10x improvement is that, as a startup, most, you're mostly odds are against your success. You might run out of cash, your talent might not come in, you might make a few mistakes that causes your company to die. So you're really prone to failure. So if the opportunity you're going after is not 10x better, the outcome is going to be 10 times better, at least, than the current outcome. Then you're, you have very limited room for mistakes because if you're 10x, the

[00:48:20] Audrow Nash: So you view it as margin in a sense,

[00:48:23] Edward Mehr: in

[00:48:24] Audrow Nash: the amount of times better. It's yeah. So if you, only do two times better, you have a very small margin

[00:48:29] Edward Mehr: yeah, you make few mistakes and now you're as good as the traditional technology, right? And then you're like, okay, is it really better? Do I want to pay more for robotics in this space?

and I think that's a problem. A lot of companies had, there are a lot of incremental improvements and I don't think you can build a company on incremental improvement, at least in this space. In finance, maybe you can, but in business space, you can't,

[00:48:51] Audrow Nash: And this space being robotics, you're saying, or technology?

[00:48:54] Edward Mehr: and slash hardware, I would say,

[00:48:56] Audrow Nash: Slash hardware. So

[00:48:58] Edward Mehr: need to have that room for, failures. And say, okay, despite all the failures, maybe I end up being 2x better. I started at 10x and then I ended up being 2x better, but still 2x better, right? so, that has been, I think, a story with a lot of robotic

[00:49:16] Audrow Nash: a really strong concept initially that creates really great improvements.

[00:49:21] Edward Mehr: You almost have to be very ambitious initially and have a path to make a very ambitious company work.

And then while you're shooting for the stars, you maybe land on the moon, right? But, but that's, I think that definitely much definitely applies in this space. And that's why we haven't seen a lot of successful companies.

[00:49:39] Audrow Nash: Interesting. What are some examples of, naming specific ones that haven't innovated, innovative, innovated, 10 times, possibly not be good, but can you give me some like cartoons of companies that have not been ambitious enough, and then I'd love to hear ones from your perspective that are.

Really pushing it, 10 times or more to be better.

[00:50:08] Edward Mehr: it's a good question. the example I give usually is, to some extent, I come from 3D printing world, and I think that could potentially be, if you consider 3D printing machines as robotic machines, they're automated

[00:50:22] Audrow Nash: Yeah, totally. They sense, they think, they act.

[00:50:24] Edward Mehr: I think that is one area where, we initially thought it was like going to be revolutionary, but I think the reach of the parts it can do is not as big as we thought.

that was the point. And if it was providing advantages in some of the applications, it was not 10 times better. Like in some applications, it was like, if it was like you're building a rocket engine, a heat exchanger additive is really good. And what if you're making a bracket for some vehicle?

Yeah. argue your way one way or another, but it's not 10 times better.

So you invest a lot of money. There's a lot of room for failure. And at the end of the day, it's not 10 times better. It doesn't drive Adoption. So I think that was one example and that's what I learned actually from, when I was in 3D printing, look at the market size of the 3D printing right now is around 16 billion, everything, plastics, metal, everything,

[00:51:17] Audrow Nash: Yeah.

[00:51:19] Edward Mehr: and there's probably close to like somewhere from 6 to 10 billion dollars in investment that went into that technology.

[00:51:26] Audrow Nash: Not that much of a return.

[00:51:27] Edward Mehr: So not, much of a return there, right?

[00:51:29] Audrow Nash: And there's a whole bunch of failures and everything too.

[00:51:31] Edward Mehr: There's a lot of failures. Yep. yeah. Now, when we started this technology, for example, like one of the things I really wanted to go after a market segment that has significant room. so that's why we went after sheet forming as a first, foray into manufacturing in this company.

and that's a 250 billion market. so lots of room for improvement, but then also, started thinking about, okay, what is the portion I'm replacing? I'm replacing the dyes, and I want to go after sector that the dyes are very, expensive, right? or the lead times are very high.

[00:52:08] Audrow Nash: Military was a great thing for that then.

[00:52:11] Edward Mehr: Yes, So, I think, yeah, 3D printing was one of those examples where, overall as an industry where the benefits were not as great as we thought. There were some applications, though. If you look at within 3D printing, you look at like dental. They found an opportunity in an application that could provide 10 times better for dentures and aligners.

They're like 3D printing can customize these things that were 10 times better than the traditional alternative, where you have to put like a little thing in someone's mouth and, and now you can scan it and you're going to get it, right? so I think that was the area that, and then you have to, mold, is expensive, you have to manufacture it, now with 3D printing, we can get aligners or dentures really, easily.

and it's also, looks much more realistic with 3D printing, you can do all kinds of colors, and you can make it really look realistic. even within 3D printing, you look at the areas where there was opportunity for improvement was 10 times, then they were successful. But any area that you were just marginal improvement, it's dying out.

[00:53:14] Audrow Nash: How would you so how did you find your area? You mentioned you wanted a big market and then you wanted you looked for where there was a lot of pain already in a sense is that kind of and then so from that you find your problem and you look in the solution space for what can be possibly solving that, and then you arrive at one that might be feasible as a 10 time solution or how do you think of it?

[00:53:39] Edward Mehr: For us it was a little bit, not as linear as you think, because I think there's three components. There needs to be a big market, there needs to be a solution, that also there needs to be enablers that you can provide to that solution, that provides that ten times increase in improvement, right? so for me, I was in additive space.

We saw the challenges we've seen with additive. So for a couple of years, actually, I started thinking about, okay, if you want autonomous and agile manufacturing that is not product specific, and it's not part specific, it's not material specific, where would that area be? I looked at post processing and manufacturing.

I looked at CNC machining. I looked at a lot of different processes and then sheet metal is the one we landed. Due to three combinations. One is, it was a huge market size. There was a good amount of legacy work that was being done. So the technology was almost ready. And then my background and my co founder's background could elevate it to the next stage, which was use robotics and artificial intelligence to bring the cost of making this manufacturing cell down, but also make it accurate, which was the only shortcoming the sheet forming had. You couldn't make accurate parts with it with incremental forming that was done in academia. so for us it was a little bit of an exploratory process. I don't think there's a, I don't think there's a recipe that I can,

[00:55:03] Audrow Nash: Yeah. It's not simple.

[00:55:03] Edward Mehr: Advertise.

You just got to iterate a bunch of processes and see where your skill set can provide an enabler, but there's still a huge market. and the jury's out. We'll see. I think we think it's a big market we are going after, but we'll see. We'll see how we do.

[00:55:19] Audrow Nash: you've made it this far, which not to say that's success is not certain by any means, but it's a good indicator I think, and especially 'cause your valuation has been good and you have, I feel like a lot of wind at your back already for a lot of manufacturing coming to the US. So yeah, jury's out, but it's looking good.

[00:55:42] Edward Mehr: Looking good so far. Yeah, I think, it's looking good so far, but it is tough. I think that's probably the toughest portion. I'm actually, I have an all hands coming up, which we're going to talk a little bit about Makida's master plan. I'm planning to actually put that on Twitter as well, or on X, so people can see.

But I think, Yeah, you need to really try to be, in early days, you need to be ambitious, but not overly ambitious in terms of tech, that the technology is not feasible. but you need to be as, as ambitious as you can think the technology can take you. So you have enough room for failure,

[00:56:17] Audrow Nash: Yeah. A way, that I've thought about it is, I've seen many companies that have ambitious plans, but their plans rely on four, five, whatever technologies that are all at the cutting edge of being ready at the same time. And that kind of thing doesn't lead to terribly good things. There's too much risk in it, in my belief.

whereas the companies that I see that are a lot more successful, they pick almost always old technology. Except for one area where they're going to innovate, and that's the one where they're going to really push hard. And they usually have some sort of like great team that is all leaders in this area.

and they're going to push that one technology to the cusp of whatever it is that they are trying to

[00:57:16] Edward Mehr: yep. Yep. It's two sides of the thing, like you need to have a big opportunity, which provides you a lot of, room to failure. But then on the risk side, if you have five or six or seven different risks in technology, then also then you're going to exhaust your ability to fail, like you're going to exhaust your failure, quota,

[00:57:35] Audrow Nash: Oh yeah.

[00:57:36] Edward Mehr: so you need to just make sure, yes, like physically you're, you, you can get there pretty fast. And this is actually feasible technology wise for sure. that's a big portion of it. Like I said, in our case it was like, robotics, can they make the price of the technology the system cheaper using robotics?

And that's why we use off the shelf robotics, as opposed to building our own gantry system or our own custom system, right?

[00:57:58] Audrow Nash: And also it's less to work on. It's one of those, you're using old established technology, not, you're not building an arm yourself.

[00:58:06] Edward Mehr: I don't need to build a tool changer for it. There are already tool changers existing for it. I don't need to build a calibration system for it.

There's already a calibration system out there. All I need to then focus on, if I use robotics off the shelf, I need to focus on my intelligence. which is enabled by these model buildings, right? So now I have one big technical risk, but it's a good risk. It's a risk that we're seeing with all the developments in AI is constantly being improved and de risked.

So you're absolutely right. You don't want to, you don't want to build a company on 10 different risks, and hope, hope that all of them will align somehow.

[00:58:42] Audrow Nash: definitely.

[00:58:43] What if you weren't venture funded?

[00:58:43] Audrow Nash: Now, not not saying accepting venture capital money is bad or anything, but I would just be curious about your thoughts on this. Do you think you could do what you've done without venture capital dollars for this kind of thing or what would the path be or how much slower would it be?

what are your thoughts around this?

[00:59:13] Edward Mehr: Venture also ties into this conversation we just had really well. Because what does Venture do? Venture is going after opportunities, the colloquially say, each check should be able to pay the fund back. So they're taking a huge risk, right? Especially in earlier stages. They're like, okay, it's 1 million check I'm writing, needs to pay this 100 million fund back.

That means there's opportunity that 1 million can turn into 100 million for every investment they make, because they probably have one percent chance of success. so by default, and some, founders sometimes complain about this and say, Oh, why is venture investing in these crazy ideas? Because that's how their model works.

Their model works, but relies on one success out of a hundred. but when you're relying on one success out of a hundred, that needs to be very big upward, very,

[01:00:07] Audrow Nash: Lots of potential.

[01:00:08] Edward Mehr: Lots of potential and very opportunistic and very huge ambitious, concept.

So ties into that conversation that we had as well, like if you're building a robotic company that's incrementally improving something, Venture wouldn't be interested in it because it'd be like, okay, how much improvement, how much room is there?

so, ties very closely. I think if you do have that concept and you're convinced yourself, very honestly that the opportunity big and I can build it The venture is good dollars. I think venture comes in they have expertise they have network. So it's not just about money It's about being able to have a network of people who can connect you to the right people.

For example, one of my investors You know connected me to a lot of folks at Stanford. I took a a program at Stanford. It's an MBA like program at Stanford. Helps a lot with building the company, hiring the right people. so I do think venture is very helpful, but not all companies are venture fundable.

if your company's not venture fundable, and you're going after venture dollars, you're in for a lot of pain. Because the expectations are going to be high. You're going to have really bad interactions with your board members.

so that's something that to be avoided. So make sure, a lot of times it might make sense to bootstrap your own company.

A lot of incremental improvements make sense to bootstrap it. try to see if you can get some private funders or your own money. Slowly build it out. but if the concept requires significant amount of capital and there's a huge opportunity, venture dollars can help a lot.

[01:01:44] Audrow Nash: Gotcha. So it's basically the upside. The potential upside is a big way to make this decision of if venture capital might be appropriate. And also, I suppose there's the kind of implicit thing of you need to have a good use for that money. but if it's Incremental, as you say, then there's probably not that good of a use for it.

You can just do it yourself. And if it is going to be a hundred X or something, then maybe venture could be a very good fit for this.

[01:02:21] Edward Mehr: Yeah, and to your point, I think one point that you brought up is pretty important to do is that raise the amount of dollars you need, because it bites you back afterwards as well, like in terms of down rounds, right? So, try to figure out how much dollar you need and try to raise as much as you need.

Not, necessarily more than that. Because that's also another challenging thing that comes with the ventures. The higher amount you raise, the higher the expectation is going to be. So make sure you're pacing your business and progress based on that.

[01:02:51] Audrow Nash: Yeah. Makes sense.

[01:02:53] Or growing quickly

[01:02:53] Audrow Nash: So you guys. You're 70 people and you've been growing really fast and you're expecting to grow a good amount more pretty quickly. how has that experience been and what have some of your lessons or findings been?

[01:03:09] Edward Mehr: Yeah, I think we're chatting this with some of the, other team members here. I think the toughest part of building a startup is actually still to this day, it is not so much the technology, it's the, people.

[01:03:22] Audrow Nash: Definitely.

[01:03:26] Edward Mehr: People are what the company is. It's not the technology.

Tomorrow, the technology can still be there, but if all those people go, we don't have much to build from. so I think the biggest lesson is hire well. and what that means is, it needs to be people who are passionate, mission driven, and excited for that technology. You don't want to just hire smart people. If you want to hire people that I have to drive, excited about the mission that you're after, the larger problem you're trying to solve, they can make light of themselves, they can, in, in, face of challenges and adverse situations, they can still move on. so there's a lot of, basically most of my advice is around.

Choose very carefully who you bring to your team, especially as you're scaling, because your company is the people you hire. So that's one condensed version of advice, but there's a lot of branches there to dive into. What does it mean to be a good employee, a good team member?

[01:04:33] Audrow Nash: Do you want to speak to it just for a bit?

[01:04:36] Edward Mehr: Yeah, I think, If I want to condense it, there's probably or two or three characteristics, that I've found is very important. Number one is mission driven. if you are building a company that provides 10 x improvement, you want people who are excited about that 10 x improvement, that vision that the, that, the company built.

You don't wanna just bring experts in robotics or machine learning that have no connection to that. they don't care if manufacturing world becomes better or worse, right? they're here to just provide their expertise. so mission drivenness is a big component of it. I think probably the most important portion. the second one is obviously you want people who are smart and that doesn't mean necessarily they they know and they're expert in that field. It's good to be expert in that field, but more important than being an expert is that somebody who can learn their way. One thing I learned throughout my career is that every few years you have to learn something new.

there's a new robotic framework, there's a new modeling technique, there's a new So yeah, and you need to spend your weekend or whatever, take a course on Coursera and learn it and see what's going on. To this date, I still do.

[01:05:52] Audrow Nash: Oh, constantly. Yeah, I'm sure.

[01:05:54] Edward Mehr: Constantly, I was like looking at the new, some new changes in transformer architecture over this weekend.

I'm just, I'm like, okay, what's going on? I had to look at it and learn it. so, that's constant. So like smart, that means that they learn fast. and that's more important than experts. and the last piece is, I think, It's, you want doers. a lot of experts and a lot of smart people would love to chat with you about how 10 million ways a thing cannot be done.

And you're like, great, I don't want to talk with you about 10 million ways that it cannot be done. what is the way that can be done? And let's try it. And the problem is in our space, there's a lot of unknown unknowns. So you can't even argue about it. You don't know, like you might think, but until you actually have done it, you don't know.

So you want doers, you want people who have more bias toward, once I have an idea, I want to get it done as soon as possible to see the results as opposed to. Stick, be stuck in analysis paralysis and chat about it for days. I think those three things, if somebody has it, I think it's a core things that, that means it's probably a good person for early stage startup.

[01:07:06] Audrow Nash: That's funny. The doer one is very funny. yeah, it's really interesting. that people can, I don't know, it's like, it's like your craftsmanship approach where you have the metal and you're gradually forming it but you're looking and seeing what you do to try to get it to the overall shape that you want.

It's like you need to do that rather than just perfect solution all at once, because that's very hard to do and come by and it might just be wrong by the time you implement it. Yeah.

[01:07:40] Edward Mehr: I mean that specific concept, when we were first starting the company, I was talking to a lot of robotic experts. And we're like, Oh, we're going to do this. And everybody was like, Oh, the robot will deflect. It's not going to be accurate. It's not going to form a part. And these are experts in robotics, like people built, segmented robots, like the ones we're using.

and then it was funny. And then six months later, we built the first system and we formed the part. And they're like, Yeah, I guess it could be done. But like the conversation before until it was done, it was all about ways that it cannot be done. and then once it's done and it's shown and it's done, then like people tear in a leaf.

but, but those are the people you don't want in your team. You almost want people who are almost like naive enough to try it. right? But they have enough expert to back and build a good system, expertise to build a good system, but they're not too smart that prevents them to actually take an action.

[01:08:31] Audrow Nash: Yeah, it's funny that we say too smart to do the thing. I don't know. It's, almost like if you're too smart you see all the things but then you can't do it. I don't think it is that though. I would imagine it's more fear of trying and failing or something like this. I think some people just don't like uncertainty and pushing into something and not knowing if it will work or not is terrifying in some sense.

[01:08:58] Edward Mehr: Yeah. I mean I have a lot of smart friends, I would say very smart, is that because they can play that chess in their head. The reason I say they're too smart is that if they play chess with me, they will always beat me because they see 10 moves ahead, right? and I almost say a person who's moving, seeing 10 moves ahead might not be a good fit for your stage startups, right?

Because, the problem is in really in chess, you might see 10 moves ahead and it's a very predictable environment. But there's a lot of unknown unknowns. so you want to almost, it's almost as explore the map. It's one of those game, strategy games where you,

[01:09:35] Audrow Nash: Fog of War,

[01:09:36] Edward Mehr: open up, let's go in that direction.

And it's the map opens up a little bit. And then we can make a decision. That's what really happens in real world. It's not really chess. So very, smart friends. And we're like, this will happen. Then this thing will happen. Then that thing happened. That's why it's not possible. And we're like, but the moment you do the first state.

There might be 10 options open up that you didn't know you couldn't think of today.

in chess, that's not the case, but in real world, that's the case. Because it's much more chaotic.

[01:10:03] Audrow Nash: totally. Or if they were playing chess, it's okay, you have to pre decide your next 10 moves or whatever, and you get to pick every single time what your move will be looking at the state of the board. You would have a pretty significant advantage .

Anyways, I wanted to get your thoughts on a bunch of like contemporary technology things.

[01:10:26] Humanoids: challenges + Elon Musk

[01:10:26] Audrow Nash: Let's start with humanoids. Tell me, about what you think, what some of the challenges will be, and then the future that you imagine.

[01:10:35] Edward Mehr: Yeah, I think humanoid's an interesting concept, I, like it. I think humanoid's from the perspective of generating excitement about robotic space is as, close as it gets in terms of oh, they look like humans. They can do everything we do.

yeah. And there are environments where you certainly need humanoids, inside homes, elderly care, might be a good, area. Even in that area, there, maybe there are other form factors that are better, but, there are certainly areas where humanoids are very good.

And like I said, PR wise, it's a very good story, because it looks like a human walking around, reasoning, talking, so it's amazing in terms of excitement. There are some challenges in terms of people are working on what does the actuator actually need to be in a humanoid? I have some of my friends working on that.

and exciting developments there. But, I think, We realized a while back that human form is not necessarily the most optimized form for all tasks. in manufacturing, we already know robotic arms are much better. They can be much more precise. They can have much more stiffness. They can apply higher loads.

So the amount of flexibility you get over the performance times performance is actually pretty big. And it's much better than a human for form factor. So I would say in areas like manufacturing. or Logistics. A simpler form might be more effective, and that mostly, I think, is robotic art, because you provide the same kinematic freedom as a human, in a much simpler, easier to manufacture robotic system.

[01:12:25] Audrow Nash: Oh, yeah,

[01:12:28] Edward Mehr: so I think they're going to be, eventually, hopefully, we're going to have a multitude of robot form factors. and it's going to be different for each application. Humanoids will have a space, but I still not, I'm still not convinced how big that space is. as, as much as people talk about it, I think it's still pretty limited because we long, too long ago, we figured out that human form is not necessarily the best form for, many applications.

[01:12:59] Audrow Nash: It's interesting. The thing that stuck out to me about your explanation is that robot arms might be the optimal form for a lot of things for what we're doing. And that's a really interesting idea. And it makes good sense to me, being that it's really just the most simple way you can have some sort of end effector moving around doing something.

So if you need.

[01:13:26] Edward Mehr: degrees of freedom on the end effector, yeah. Yep.

[01:13:29] Audrow Nash: So it's the simplest way for that, and also if you look of robotics at the moment is that a lot of the things that are very valuable that we're doing are simple mobile robots, where you're just having a robot drive something around moving from here to there because we have a lot of those types of problems and the state of the technology is very well suited for that type of problem at the moment, and what I would bet, connecting with your idea, is that the next like 10 years or something might be where we see a lot of robotic arms doing useful things in more spaces.

but, and they'll probably be a more optimal form factor than a humanoid which has a smaller arm, or smaller arms, not as much reach, not as much torque, or stiffness, or these kinds of things. I wonder, what do you think of that idea?

[01:14:31] Edward Mehr: No, I think, you're absolutely right. I think there's analogies in other industries as well. I think of robotic arms, they first got deployed in 40s and 30s, roughly around that time. I don't know exactly when was the first robotic arm was made, but it was somewhere in Germany was used for, in 30s or something or 20s.

so it has been in the industry for a long time. So there's a lot of legacy has been perfected for a long time. Now there's multiple vendors that are competing, so it's almost commoditized. so it's a very good interface. The parallel of this is to, we see a lot of growth in NVIDIA, growth in NVIDIA today.

What was that company? It was a

[01:15:12] Audrow Nash: Just an enabler. Yeah.

[01:15:14] Edward Mehr: Gaming chip company, right?

[01:15:16] Audrow Nash: Uh-huh..

[01:15:17] Edward Mehr: There is something to be said why NVIDIA became cornerstone of, computation for neural networks. It's because of the legacy. It's a chip that already existed maybe for another application, but it's a very good fit. and it was perfected.

Now, maybe people are thinking about other types of chips. But the foundation was built on a legacy system that had other applications. Because all the bugs, all the issues have been ironed out, and it's easy to use it, and it's commoditized, and the price is low, and you can just use it.

I think the same analogy works for robotic arms, where these systems have been perfected.

We have a very good supply chain, Siemens makes a shit ton of, electric motors, and Nabtesco makes drives, and they have perfected this supply chain. why do we want to disrupt it for another form factor? There needs to be very good reasons, and I think that the established supply chain usually wins out if the improvements are not ten times better.

[01:16:24] Audrow Nash: because they get a lot of efficiency from the amount of time they've been in the space and the amount products that they've pushed out. Yeah, that's a very cool perspective. I'll have to think more about that.

I like that a lot. let's see, do you, what, so one thing that I think is especially interesting with humanoids is you have large valuations, like Figure, I don't know, raising a billion dollars or whatever it is. and they are their own company raising the money. I wonder if it will be like autonomous cars, where you see Cruise gets a billion dollars and then its very hard to deliver on the product in the long run. But the thing that's especially interesting to me is then Tesla with their Tesla bot, or Optimus, and the thing that's really cool with that is that they are their own captive, market in a sense.

Like it's a very farsighted initiative. They don't have to appease venture capitalists, who want to return on investment within some timeframe of years. so I feel like it's an interesting dynamic, that they have been able to fund this research internally and maybe in on a longer time horizon it will be exciting or maybe just the research can be applied to other areas like maybe the top half of a humanoid is very good for upholstery in car manufacturing for example.

What are your thoughts on this?

[01:18:07] Edward Mehr: Yeah, it's a good point you brought up. I think all Elon companies should be seen in a specific light.

[01:18:16] Audrow Nash: I'm excited to hear what that light is. Yeah, go ahead.

[01:18:19] Edward Mehr: I think one portion of it is like he does things that are very exciting, right? Now he's, in charge of this platform X. And, if you look at a lot of the companies, he's, he goes after that PR angle, right?

because I think that's what, That's what generates enough excitement for him to create a big problem space that he can solve. But the angle that people usually forget, I think, about Elon is that he does everything, in my opinion, and I may be wrong, but based on my experience working at SpaceX and interacting with a lot of people who work with Elon even today, Elon's number one goal always has been becoming multi planetary, and I think you can see most of his companies in light of that.

Obviously SpaceX is at the core of it, with the Starship and Falcon 9, finding a way where we can get to Mars and other planets. Electric propulsion is also another reason that, that's pretty much the only way you can travel in other places. You cannot use, necessarily a biofuels and things that we have been

[01:19:24] Audrow Nash: You run out or

[01:19:26] Edward Mehr: out. Yes, electric is probably the most, available way of doing propulsion in those places. and then you look at Boring Company. If we go to Mars, there's going to be radiation out there, we need to go down in the ground. So

[01:19:40] Audrow Nash: Aha, that's wild. I never connected all this. Okay.

[01:19:44] Edward Mehr: you need to go down and do that. neural link is also, we're going to come back to it, but it's, about kind of high bandwidth control interface.

and then I think of humanoids also an easy interface for astronauts. if you build a system for humans, then you can replace them with humanoids for, outer space applications. So a lot of times, a lot of bend to Elon's companies are around, I think he's building different components of that larger term goal that he has, which is becoming multi planetary.

And he's being very creative about it, like finding business use cases on earth that can propel this. But I do think he always had that bend toward multi planetary and that's why he starts from there basically. He looks at what do I need on Mars and then

[01:20:32] Audrow Nash: aha.

big build backwards and what do you need to build here on Earth.

So that's that Figure. I don't know. I don't know enough to know, but I think Elon companies at least are unique in that perspective.

Yeah, that is a very interesting way to frame it. And it makes sense. I had never connected those dots. I love it. Hell yeah.

[01:20:54] AI trends: transformers + hallucinations

[01:20:54] Audrow Nash: And then so now moving on to AI. when you guys, when you, when we started, you were saying AI, you're using AI and I think AI has become a good bit of a like buzzword marketing thing.

Like everyone wants to say AI when they mean machine learning,

[01:21:09] Edward Mehr: Yeah.

[01:21:10] Audrow Nash: which is optimization from my perspective, like regression or things like this. And, so from my perspective of what you guys are doing, it's not chatGPT, it's not, you're using neural networks, but I, don't know, tell me about your thoughts on AI and then we'll go into more specifics.

[01:21:38] Edward Mehr: yeah. It's interesting. the definition of AI has changed

[01:21:43] Audrow Nash: Oh yeah,

[01:21:44] Edward Mehr: over time, right? I think, like I said, if you look at back in 60s and 70s, like rule based systems were considered artificial intelligence. then when I was in school in 2004, we were talking about kind of tree search as a AI system where like you basically, like rolling out future in multiple steps, you optimize and then you get that.

And then with the data, the definition moved toward ML. Basically, can I find the pattern? In, in a set of data, the same way human finds a pattern, like you look at a picture and the pattern of colors and lines tells you it's a dog. and that became like the state of the art for AI. Now, I think with transformers, that again, change into large language models, became the state of the art of AI because you could talk to it and it will talk to you like, like a human.

But the core of large language models that are GPT or transformers, is the same as, to me, for me, it's the same as a more simpler neural network, but, and then

[01:22:52] Audrow Nash: It just

[01:22:52] Edward Mehr: is the same as, and each node of those is basically a regression mode.

[01:22:56] Audrow Nash: Yeah,

[01:22:58] Edward Mehr: so they're all almost like maturity of the same system, all the way even to large language models. so I categorize them as the same technology, and AI being the term that we always apply to the most advanced cutting edge portion of that. so yes, we do use machine learning. we are using graph, neural networks. So closer, and then with this next project, we are using transformers, to do a multimodal,

of

[01:23:31] Audrow Nash: So you can have a context remembered

[01:23:33] Edward Mehr: You can have a context and you can also do, we're talking about multimodal models now where you can feed it an image of a CAD of a part, but you can also feed it an instruction of what accuracy you want. And then come up with the robot actions. So it's a multimodal model that we are looking at

[01:23:48] Audrow Nash: That's super cool. Okay, I got to learn more about that kind of thing.

[01:23:52] Edward Mehr: So that is closer to what is happening with ChatGPT. Now, I think the definition of the AI will progress every time we beat the boundary. Now, is it the Terminator or the, Skynet? Not yet. in pop culture, that's what people think of AI. But, if you go into academic text, the definition has changed over time.

So I don't know, I don't know what to call actually. like you can talk about, go with a pop culture reference. I don't think anybody's there yet. but if you go with the LLM, then you can talk about, okay, transformers are maybe AI, building components AI. We are modeling using transformers. So it's a question of how you define it.

[01:24:37] Audrow Nash: yeah, for sure. Yeah, it's been funny that it's whatever is, I don't know, I guess AGI Artificial General Intelligence is like the new thing everyone's trying to do and that's the holy grail of all this AI stuff but it's, funny that it's oh, artificial intelligence will never get this and then it completes in chess and then it.

competes and beats the, top guy or the top person and go, and then it's oh, it's not AI yet. Every, one of these ones we redefine ai. yeah. I see what you mean. Where we change the definition. What are your thoughts on ChatGPT, or large language models more specifically and where they're useful, I suppose?

[01:25:27] Edward Mehr: Yeah, so I think The way I think about large language models, and I'm not an expert in the field, I have very surface knowledge, in that field. I think there's something special about them. obviously they're language models, so they can just talk, right? But I think they're capturing the way we think, because if you can say, I think of language as the way we are encoding a human knowledge and thought process, I think mastering that is as close as possible we have figured out to this date into, into, thinking.

So then there are areas in, manufacturing where that directly applies, right? and mostly is around interface improvement. if I'm programming a robot to do certain things, can I just talk to it? I can just get to a point where I say, Okay, should I code up forming this geometry?

Or can I just say, form this CAD and then say, Okay. Scan it and they will tell me what is the metric and I say can you improve it by 10 percent and then they will do another round of forming. So the direct area there is like removing friction in an interface with robotics. And I think that definitely is on our horizon as well.

We're looking into that. Maybe not first order problems we want to solve but it's definitely in our horizon.

[01:26:44] Audrow Nash: Eventually. Yeah.

[01:26:46] Edward Mehr: But with multimodal models I think we are getting, much closer to that being applied in manufacturing. I was following one of our great friends, Pieter Abbeel, at Covariant, they released RFM1 where it's a multimodal model.

You can just say, okay, here's a picture of a bin and here's an instruction, pick up a banana from it. And the robot will, the model outputs robot actions on the other end. We'll pick up a banana from that image. so now. Now we're getting not just text, we're using the same component, which are transformers, same building block, but creating these multi modal, multi interface, multi medium, models that can input image, still frame, large language, and then outputting the same thing.

[01:27:36] Audrow Nash: You just convert it to data and it will do something with the data.

[01:27:38] Edward Mehr: And they're combining image and natural language in a way that, the brain does. and that's pretty exciting. I think that has a huge potential to interrupt, disrupt, the robotic space.

[01:27:51] Audrow Nash: Gotcha. One of the things that was interesting to me about especially working with ChatGPT and maybe this won't be as much of an issue for robotics companies because the task space could be smaller but there's a lot of weirdness around predictable results where it'll just hallucinate something and they are getting better like specifically in my experience lately, Claude3 is way better, at hallucinating far less than ChatGPT.

That has been really cool. That actually might be the first one that I can actually use rather than just getting really early ideas about something or like very basic coding help. that one might actually be, I might actually be able to enlist it for like podcast related tasks. But, what do you think about the hallucinations of these models?

[01:28:48] Edward Mehr: Yeah, this is a new, not a new thing, right? we're from a robotic system. We always, we've worked with deterministic systems where same input gives you the same output every time. But we have been dealing with humans and humans are very non deterministic, right?

And they are a big part of manufacturing process, right? Humans are a big part of manufacturing and the way we got around that is through checks and balances. Right, and we say, okay, if the output of this result needs to verify this way, as long as it's passing these verifications is good enough. Because in reality, we could not rely on human output.

One day I'm sick and my, mistake, make a mistake. And the next day I'm, I am, I might outperform what I usually do, right? So we already have built systems that rely with, that, deals with ambiguity. So I'm not that concerned about it, honestly.

[01:29:39] Audrow Nash: Yeah you just have your checks and verifications I think that's great because you're right, you don't need to give it total control and have it do crazy. if you were having a robot choose actions from a language model or maybe like a language action model or something like this, or a large action model, you could just run it through a path planner at a very high level with a very low fidelity simulation really quickly to make sure it's not going to collide with stuff.

And that would be a very simple check on the proposed actions, which is actually a very cool idea.

[01:30:11] Edward Mehr: No, you're absolutely right. And then the other piece is, I think, at least the way we thought about it, is that once you want repeatability. You can use neural networks and machine learning and AI to get, to develop something and say, okay, this is the, like in our case, for example, you can ask a model to just say, what is the process parameters?

What path the robot needs to take to form this part, this aircraft wing, below one millimeter accuracy. And it will come up with that. And maybe you make a mistake and you have to fine tune the model. But then the moment you want to make 10 of them, then close it. Don't ask every time from chat, from your model to how to make the same thing over and over

[01:30:52] Audrow Nash: You already got

[01:30:53] Edward Mehr: close it.

So you can also create these, almost silos and be like, okay, in development, I'll use the models then, but in production, I'll lock it down. at least that's how we were thinking about it. For example, for our case is that we don't want, once we figured out how to form this wing, we don't want to ask the AI anymore.

We just want to replicate the recipe it came up with.

[01:31:15] Audrow Nash: Yeah. I like that a lot. Any, thoughts on the form that these checks and verifications for LLMs, like I'm just thinking of like unit and integration testing for this kind of thing, and maybe some like high level safeguards, but any thoughts on this? Cause it's a really cool idea. This is what I'm going to think a lot more about, I think.

[01:31:35] Edward Mehr: Yeah, I think you nailed it in most respects. As long as it's not going to destroy the system, I think you can let the robot do what it needs to do. As long as it's not destroying itself. it's that means, maybe don't go above certain currents, don't run into yourself.

You just talked about collision.

[01:31:51] Audrow Nash: So you give it a, safe like parameter, a safe set of actions,

[01:31:57] Edward Mehr: yes, and that's a much bigger space than what you could think of, but at least the robot doesn't damage itself.

[01:32:03] Audrow Nash: Yeah.

[01:32:04] Edward Mehr: And I think as long as it's not damaging itself, I think you can probably let it explore.

[01:32:11] Audrow Nash: huh.

[01:32:12] Edward Mehr: At least that's how I'm thinking about it.

[01:32:14] Audrow Nash: Yeah, that makes a lot of sense to me. What do you think about the role of simulation? In all of this and sim2real and everything like this

[01:32:22] Edward Mehr: Yeah. That's another interesting area, right? like you can let these robots to be in the simulation world. For us, it's a little bit tougher because the main reason we went to AI based models is that physics simulation was complicated, and tough to do. but that's another branch of work we're doing at the moment.

we actually did this proposal with, Lawrence Livermore at the National Lab a while back and trying to create a simulation of our process that is on GPUs and it's faster. so I think there's opportunity there to create augmented data, even if, as long as the simulation is fast, maybe it's not as accurate, but if it's fast, then I think it can be helpful, right?

Because you get like roughly the right thing, and then you can do a lot of checks and balances, and it generates you a lot of synthetic data to also improve the model. As long as directionally, it's correct. If the model is directionally not correct, then it's going to be really tough.

[01:33:23] Audrow Nash: what do you mean directionally if it's doing completely the wrong thing

[01:33:26] Edward Mehr: yeah, completely wrong.

It's maybe it doesn't accurately figure out what's going to happen in the real world, but you know that if you increase the speed, it will, the outcome also moves in the right direction. That is similar to what happens in the real world.

[01:33:42] Audrow Nash: Yeah, one work that I saw quite a while ago I'm sure you guys are aware of stuff like this, but they were using a very high fidelity simulator and they were using machine learning to try to link one frame to the next of complex objects. And so the simulator, they ran it super slow in super high resolution, very tiny timestamp, all these things like this.

But they learned the transitions and then within a space of. Things that it had seen it was a very fast simulator because you just had a pass through the feed forward network effectively. and that kind of thing seems really cool. And if you apply a check on that, which says, Hey, are you would like super low resolution version of the simulation?

Let me make sure that things are not just completely wonky, everything's shooting across the screen or whatever. Then you could probably use that kind of thing for learning, at a fast rate, which is quite cool.

[01:34:44] Edward Mehr: No, I agree. I think that, that building that kind of a proxy. you're almost building an empirical machine learning based proxy model for the simulation using that technique.

which is, fantastic. I think that's a good way of, if you cannot make a lot of data in the real world, use the simulation to augment it.

And as long as, like you said, directionally is correct. It's not going to do, if I move this to the right in the simulation, it's going to move right in the real world as well. As long as that directionality is there, I think you can generate a lot of data and learn a lot of patterns from it for your model, and then retrain it or do transfer learning in the real world to then fine tune it based on what's happening in the real world.

[01:35:27] Audrow Nash: Oh, yeah. changing topic a little bit.

[01:35:32] Thoughts on manufacturing

[01:35:32] Audrow Nash: Looking out at the manufacturing space? What are some things that are top of mind for you, predictions maybe in the next five years or so? or where are you watching? what's interesting to you, looking out at the future and considering our world and manufacturing?

[01:35:51] Edward Mehr: Yeah, like we said, there's a lot of wind for, on shoring manufacturing. I think there's going to be a lot of policy, both investment and the government investment and private investment in that space. But what I think is, it's interesting, there's two camps. There's a camp of Let's replicate what we did in the 1960s and 1940s and World War II.

Let's bring all that scale back. oh, let's create foundries, and let's create, machine shops, and let's make a lot of them, right? And there is, I think, a camp of Let's reinvent those technologies so that they can actually be economical in the United States.

[01:36:34] Audrow Nash: Totally.

[01:36:35] Edward Mehr: and I think it's interesting.

I think the first cap is easy to argue for. It's, I think to me, it's a lazy idea. and there's a lot of tractions for it, where we're just saying, let's, do what we did in the 1940s. and 1930s. Let's just scale it. But there's a reason we lost it. there's a reason that as the, wages and standard living increased in the United States, we offshored these things to other places.

That reason is not gone, right? So I think the manufacturer will not look like, the future manufacturer will not look like 1940s. We're not going to be able to replicate what happens in China here.

[01:37:16] Audrow Nash: And

[01:37:16] Edward Mehr: you want to beat China,

[01:37:18] Audrow Nash: it doesn't make any sense here. our populations are, we're looking like there's fewer younger people and they all want to do jobs that are not manual. So it's

[01:37:30] Edward Mehr: But there's a lot of push. A lot of people are pushing towards okay, let's bring these capabilities back. So we'll see if that pans out. I think jurys still out. If that's, in my opinion, that's not going to happen, or it will be a lot of failed

[01:37:46] Audrow Nash: option where we just try to bring everything

[01:37:48] Edward Mehr: bring back and they're going to

[01:37:49] Audrow Nash: have the people for it. I think

[01:37:52] Edward Mehr: but they're going to push forward.

I think politicians will push forward. A lot of people will push forward, oh, let's make thousand factories here that are same old technologies, right? I do think though, We need to be a little bit smarter about it and take the harder route and create new manufacturing technologies that give us a competitive advantage versus the traditional that our opponents are using, because that's how we're going to have a sustained advantage.

same cat foundry will not work. It's a more advanced foundry here is going to work. That's going to be robot powered and it's going to be, intelligent. Same sheet forming and stamping in China is not going to work here. We need to figure out how we can create technologies that form sheet metal in a much more different way, in a much more agile way.

Gives you a more significant advantage over the current technology, not just scaling it with dollars. So I think that's basically the direction I'm predicting. I think both will happen. One of them will likely fail, and it's not going to be as fruitful. and then the other, I think, is the route to go.

The latter is the route to go.

[01:39:02] Audrow Nash: I think with COVID, and Like maybe lockdowns in China or something. The, like a lot of textiles were not coming to the US and so I think we innovated quite a bit. I heard this somewhere, and I am not a hundred percent sure, but what I ha, what I believe I heard was that the, Textile manufacturing had a big renaissance here because we evolved or we created new ways of doing the manufacturing so we could have the stuff we wanted. and if say, transport lines and if the, if it's harder to get stuff from abroad, Hopefully a lot of the things we can figure out how to make here in the US out of ingenuity, like that story about the fabrics and what you guys are doing, making it more agile, manufacturing, using our strengths, which is technology, I would say.

[01:40:04] Edward Mehr: yes, they're similar thinking is also existent. in military and to some extent some of it is good like for example we were talking about okay you know what China can have higher production rate why can't we just manufacture drones that are maybe lower quality but the same production rate right and and there's no type of thinking moves you toward maybe we need a little bit of it but moves you toward replicate what's happening in China which i don't think it was like at best we're going to become as good as China and then But then we don't have the demographic, the workforce to back it up.

So let's focus on it like a different way of making. Maybe we should make drones that are, have the same agility, same production rate, but maybe last longer. They're better. and let's invest in those technologies. So anyway, I think there is an easy way out, which is like replicating what we had in the past, which might

[01:40:58] Audrow Nash: But may not work for this kind of thing, as you're suggesting. Yeah, and then there's the other way, which is make better ways of doing things. And that's probably the best way to do it. I forget who I was talking to, but someone was saying that, if we wanted to build an iPhone in the US, they have and take all of this with salt, because I don't remember the specifics but it's like there's a whole city in China that has like a million people that all work for the iPhone manufacturers and like we don't have imagine a full city that is a hundred percent around the fabrication of an iPhone. Which is bonkers to consider here. Like I'm in San Antonio, Texas and it is a big city, but we have so many industries here.

It's not like we don't just all work on an iPhone for this kind of thing or

[01:41:52] Edward Mehr: no, you're absolutely right. you look at

[01:41:53] Audrow Nash: to live.

[01:41:55] Edward Mehr: it's machine, operators running these machines over and over, again, making phones and things like that. But yeah, I don't think it will work here. there's a reason, worked in 1940s, but I don't think it worked now.

so yes, I'm in, in total agreement with you.

[01:42:12] Audrow Nash: Hell yeah.

[01:42:15] Advice to get involved in manufacturing

[01:42:15] Audrow Nash: What advice do you have? So say someone is in the early part of their career and they want to get more involved in the manufacturing space. how would they do so? they have a background in technology and they want to get more involved in the manufacturing space. what would, what advice could grease their path, make it a bit easier to get started and productively contributing?

[01:42:43] Edward Mehr: I think, obviously there's like millions of ways to get involved, so whatever I'm suggesting is probably like something that I anchored on

[01:42:51] Audrow Nash: Yeah, for sure.

[01:42:51] Edward Mehr: more relevant to me. I think the big, we have a lot of smart people working on a lot of smart things, right? folks in AI and software, electrical engineering, a lot of folks doing a lot of smart things. I have noticed that people think of manufacturing out of reach, because as a society, we don't grow up around manufacturing anymore, right? We grow up in a lot of technology, but we don't see manufacturing day to day, and to some extent, that removed that agency and that removed that believe that, I can do this, right?

one of the things that happened to me when I went at SpaceX, I used to do a lot of manufacturing as a kid, but when I went out to SpaceX, the large scale of operation, the manufacturing operations we were running, gave me and a lot of other people who were at SpaceX a lot of confidence. They thought, oh, it's totally fine, I can do this, right?

So what I would suggest to people who are smart and have been working in other disciplines that are much more theoretical, you're a smart guy, just start making something.

learn to weld, go learn to do a carpentry, do some kind of a fabrication, and maybe get involved with people who do fabrication.

But I think the biggest portion is once you do it, you realize, oh, first of all, it's really fun. Second of all. I can do it. Now I can apply this skill set that I had in robotics to do it as well. when I was doing sheet shaping by hand, I was like, okay, I can do it sheet by sheet, I can make a door for a car, whatever.

But then I was like, what, I spend a lot of time in robotic and computer engineering. Maybe I can do this. I can, make it, make, put a robot together and build this. I think creativity comes from when you connect multiple disciplines. So go do those disciplines. Don't be scared of going out there and taking a welding class.

Don't be scared of go taking a carpentry class. And then start thinking about how you can connect the dots between multiple disciplines that you might be involved. and I think that's how we're going to build those next generation manufacturing technologies that you talked about and not just replicate the past.

Is these smart people coming from world of physics and electrical engineering and robotics and AI doing these old mundane tasks that sound like mundane, but then you're like, Oh no, like I can significantly drive optimizations here and improve these techniques by 10X, 20X, 30X.

does that, make sense?

[01:45:19] Audrow Nash: Totally get involved, start building stuff, go do a welding class, do a carpenter class. I think I'm gonna follow your advice with that. 'cause I am, coding all day and I love it. But I also would love to do more with my hands, and building stuff. I'm like, I, have a pretty decent background of building things around, but I don't do much lately and I feel like I'm losing it for this, like simple robotics projects have been challenging.

[01:45:49] Edward Mehr: Yeah. And I think the good thing about it is it's funny, even when you talk about the craftsmen, you start having these conversations of material and properties. It doesn't sound like a material scientist, but it's so much deep knowledge that you gather by, by just talking to physics world, the world, the physics of the world becomes more accessible to you and you can reason it with it in intuitive ways.

When you start building versus before, it's just like textbook and Oh, elastic versus plastic strain, stress strain curve. Like you really get it. And then you start like feeling it with your hands, what that means. And it allows you to have very intelligent conversations and develop very complex topics.

Much simpler, as opposed to coming from a metallurgy school or whatever. Anyway, I think it's, it's hard to describe it, but I think once you start putting your hands on things and building things, somehow you feel much more empowered, to do things.

[01:46:49] Audrow Nash: Hell yeah. Okay, we'll end there. Great conversation, Ed. Really appreciate you being on the podcast.

[01:46:57] Edward Mehr: same, thanks for, for the engaging conversation and, I hope other people's enjoyed as well. I definitely did.

[01:47:04] Audrow Nash: Bye, everyone.

That's it. You made it.

What did you think? Is Machina Labs onto something? Do you see more great uses for AI in robotics in manufacturing? Let me know in the comments or on X. See you next time.