TRANSCRIPT 🔗
Carrie Lee: Welcome to Vision Forward. This is Appara AI's podcast for manufacturing leaders navigating the AI revolution. Each episode will feature candid conversations with Appara internal experts, our customer champions, and industry voices. We'll be exploring what's working, what's next, and what it really takes to build smart manufacturing operations, leveraging the greatest and latest in AI. This is our internal inaugural podcast. My name is Peter Brouer. I'm responsible for the sales and marketing organization at opera. And my guest today is Carrie Lee. Carrie's responsible for our sales organization across the Americas as a sales director. And our inaugural guest is Jamie Westall. Jamie is a director of engineering for Apera and is one of the early pioneers and architect of the opera technology. Thanks, Peter. So ⁓ to start off today, Jamie, why don't you tell us a little bit about who you are and what you do and what brought you to this world of AI and manufacturing? Sure, thanks for having me. ⁓ as Peter mentioned, I lead the engineering organization here at Appara AI. We have roughly 30 engineers building the vision software and the AI behind the scenes. And we're based out of Vancouver, BC, Canada. I've been with the company for coming up on I between five and six years now. And what really brought me into this team was first of all, ⁓ it's a group of really smart people. And when you have a group of smart people in the same room, magic happens. And that's what I saw happening early on. There's this opportunity to do something different, to ⁓ to to bring AI, bring innovation to an industry that was ripe for it. So that was the early days of Opera. I've it's been so exciting to be here for from when it was just a handful of people all the way to today, which is you know a world class organization installing AI vision in in in industries across the world. ⁓ it it's been amazing so far. Fantastic. Jamie before we go Deep dive the technology and commercial journey of a para. Let's take a step back. So just so tell us, how would you spend your ideal Saturday? ⁓ that's a good question. I I think you find find out a lot about someone ⁓ by understanding what they do in their spare time. for me, maybe it's gonna sound a bit boring, but my my my ideal Saturday is just to wake up, have a great cup of coffee. I I I self-identify as a coffee snob here in Vancouver. So we have a lot lots of different ⁓ coffee options, but definitely coffee first. before too long, my dog is bugging me to go out and that's actually a nice part of my Saturday. It's you know, we're we're treated with a lot of nature and hikes and any anything ⁓ to get outside in Vancouver here. So I'll take my dog on a big long walk. We're maybe at a park, maybe Pacific Spirit Park here in Vancouver, Stanley Park. ⁓ and I say it's for him, but you know, it's really for me. It's a nice long relaxation, decompress after a busy week. and and and just you know, get that get that oxygen from the nature. ⁓ after that, maybe it's another coffee. Hey, it's Saturday. but then you know, it's pretty chill for the rest of the day. I think there's Gonna be probably a workout in there and you know, cycling around the seawall maybe. ⁓ and then end it off with just a nice a nice dinner and a bottle of wine. What what more can you ask for? That that that's the ideal Saturday for me. That sounds great. What kind of dog do you have? It's a Swiss mountain dog. It's a big boy, it's hundred and twenty pounds, so he he rules the roost around here. Well, you definitely live in the best city, Vancouver, to spend the ideal Saturday. That's right. Can't complain. Okay, so ⁓ we've got a lot of questions and things we want to cover with you today, but before we get too far into it, you know, Jamie, you mentioned you're one of the earliest engineers at Apera. ⁓ you've kind of seen the company grow. But for someone, you know, this is our inaugural podcast, for someone who's listening who might not be too familiar with Apera, how would you describe, you know, what we do, who we are, and why does it matter right now, Jamie? Yeah, for sure. ⁓ so Appara AI develops vision software that that lets industrial robots see and handle parts in the world just like humans would. Okay, let's go one level deeper. What does that really mean? the Appara software has two components. One is ApparaView. This is the runtime, this is the software that runs right next to a robot on an industrial factory production line. ⁓ This is connecting to cameras, seeing the world that the ro that the robot is living in, and detecting objects in the scene and under and sending up motion commands to the robot to be able to navigate the world and pick ⁓ a a part so that it could then place it onto the production line. For example, ⁓ the view is the runtime. View is what's running the AI models on the vision. That is captured from the cameras and telling the robot where to move. What else is in Appara's software stack? We have Appara Forge. This has two use cases. This is hosted in the web, so you can go to forge.appera.ai. And first of all, you want to build your simulation. You before you even design your cell, before you buy a robot, before you buy an end-of-arm tooling, Forge is a place where you can simulate. Say, is this going to work? Will my robot actually reach the bottom of the bin or will it be able to pick this part? Will this tool work? let's run a simulation to get that confidence. And then, you know, the second use case or the the connective tissue here is once that is completed and you have some confidence that this is gonna work, you just press the button for submit AI training. Six, twelve hours later, that AO those AI models are trained. and deployed to that edge device so that view a pair of view can can run at the edge next to the robot and perform magic. It really it really seems like magic. I know I'm gonna I'm gonna reuse that word too much, but once you see it, it's
Load full transcript
it you can't believe it. It is detecting these parts that you know maybe it's completely covered, but maybe it's it's only 25% of the part, 10% of the part is showing, but the vision has learned the AI training has learned that that no that's a part Just like a human would say, ⁓ yeah, that's a part. Something's covering it, but that's a part. ⁓ and so that's that's a pair-of-view software. And you know, who do we serve? What who who's benefiting from this? ⁓ primarily manufacturers. ⁓ we're heaviest in automation, in automotive, sorry. tier one automotives, you know, these why is that? This is where the robots are in the world, this is where the industrial robots are, this is where competency and sales and robots have been installed, and they're and People are looking for improvements. They're looking for ⁓ automation. They're looking for how can we do more with these robots that we have? ⁓ so that's where a lot of our customers are. ⁓ and and you t you asked an interesting question there. Well, why it matters? ⁓ why it matters now. You know, right now, or as we've seen over the last several years, ⁓ The cost of labor and the availability of labor is the the cost of labor is going up and the availability is i it's hard to find. And so when we talk to a customer and say, Hey, you this can be done o automatically. This can be done you can have one person managing three cells of robots rather than humans, you know, picking parts out of bins and putting them onto the production line. ⁓ you know, their eyes light up. Is it this We can't find enough people, so this is great. So that's why it matters now. That's why we keep feeling that demand from the industry to do more. ⁓ and and that's what really guides our product roadmap, really. Awesome. Thank you, Jamie. So in the industry, there's there's a perceived gap at times between what an AI vision system can do in the simulated environment versus what happens in the production on the floor in the production plant. In your experience, how does a para bridge that gap? That's a that's a great question because we come up against that a lot. You know, it is one thing to show someone the simulation and then say this is what's gonna happen, trust us. But it's another to really gain that conviction that this is going to work. Specifically for a para AI's approach to AI vision, we you know that that one click button that I that I talked about in Forge to submit the AI training. That kicks off an AI training that's completely trained on synthetically generated images so that the AI model can learn what what these parts look like, what the environment looks like, what the robot is dealing with in the scene. Okay, but how how does it do that and when what's the benefit of that? And will it really work? You know, it's comparing to classical approaches where Maybe the term template matching is used. You s you show the actual part to a vision system and say, This is the part, please find it. Well, you're kind of hoping that all the other parts look exactly like that part, hoping that the other parts don't have scratches or don't have parts missing or don't have or or that the lighting doesn't change. Now, when you train with synthetically generated images, synthetic data. You can actually vary all those components and say this is a part. Now let me put a scratch on it. This is also a part. Let me dim the lights. This is a part. And this isn't humans doing this. This is automation. This is computers. This is cloud compute doing this millions and millions of times. So that just like humans learn over time, ⁓ yeah, I've seen a million of these. That's what that's what we're doing. It's it it's clo it close that gap to. What is actually happening in the world by just simulating it millions of times, generating millions of images so that the AI model learns what it's up against. It's really impressive when you put it that way. Think about the sheer quantity of iterations that we're able to go through. It's it's really exciting. a little bit ago you mentioned that, you know, a lot of opera AI systems are deployed at in automotive, you know, we're in some of the top OEMs and tier suppliers here in North America. I'm curious from your standpoint as an engineer, what makes automotive such a demanding proving ground? And What his you know, kind of building technology and solutions for that environment taught you and your team about real what reliability really has to be. This is important and and this is a key differentiator, a a key fact about our industry, about what a para AI delivers. ⁓ automotive is brutal. It it you need to be a hundred percent. These production lines cannot go down. If they're d down for a minute, they're losing thousands of dollars, these companies. That's you know, compared to other use cases of AI where You know, if someone's getting an AI model 70% correct, they're happy with that. Hey, we can't we can't afford that. We need to be n close to 100%, adding nines on the 99.9 as much as we can. And then on top of that, you know, that's just the vision aspect of it. There's robot guidance, there's how we're moving the robot, and what are the fail-safes? Okay, how do we protect? the ro ⁓ you know, if there if something goes awry, if something enters the cell, how do we detect that? How do we avoid crashes of the robot? It just cannot be tolerated. So ⁓ that is really the you know, the the the proving ground of automotive and although I say it's brutal and it's tough, ⁓ it it's it shows who's who's g it it's really tough to to execute and it shows the best of the best can survive. And so that's what we hang our hat on that. We say, yeah, we have our software running in Tier 1 automotives and it's keeping the line up perpetually. And you know, our support line is sometimes like the Maytag repair man for those people that remember that reference. It's you we don't get caught. It does it just runs and we don't have to worry about it. So we really
hang our pet on the reliability that it takes to be in the automotive industry. It's impressive. Very exciting. Hey Jamie. Often myself or my sales team will approach you and your team and say, hey, there is another application we need to solve for this customer to be able to automate a certain application. So in your experience, what what's the which application or customer challenge seems to keep you up the most at night? And why why trying to solve it will help dictate where our technologies had. Yes, yeah. There's a there's been a few. There's been a few, certainly. one that one that sticks out is maybe ⁓ a a case where we're we've delivered a solution and we've understood the requirements and it's perf it it it's being deployed on the production line, but there's a There's a change in requirements, there's a change in the environment. There's inevitably deployments go a bit different from what was planned. And we're watching this vision system get deployed. Everything's working as we expect. These dowel pins, we're guiding the robot's dowel pins into these parts, into the holes of these stamp sheet metal parts. I think there's sides of car doors. And there's just scraping the side of the hole just a little bit. And you know, ostensibly it looks like well, that's what a dowel pin is for. They're chamfered, they're just gonna find their way into the hole. This work we're hitting our one millimeter accuracy requirement. And the request comes from the customer. Hmm, we can't really have that scraping on the side. ⁓ okay, okay. We're gonna have to eke out some more accuracy here. Is it possible we're really We're really at the limits here, ⁓ of of what we can imagine. We look we talk to the customer, we look at it, we look at it internally, we brainstorm and we say, you know what, this can be done. We we have the technology, we have the ability to do this, let's optimize our training process for this customer using the samples of what we see here. we have these, you know, our technology relies on two high resolution 2D cameras and we have the resolution. Okay. This keep this this my team is working through the night. We're you know staying up, we're we're focusing on how can we do this for these specific applications that need sub millimeter. We start calling it sub sub millimeter because this has got to be perfect. And we do another iteration of synthetic data generation and AI training. We deploy to the to the to the facility and we give it another go. And I just remember telling my team, I'm I'm on site watching this because this is kind of high profile. I'm there, and we're kind of look like rubber hits the road here. We're watching it. And I just say to my team, swish, swish, swish. It's just the dial pins going through these, these, these holes on the stamp sheet metal parts, just like nothing. Just air on either side. And this is like minute ⁓ tolerances. And we test out the rest of the the the models on the rest of the robots and everything goes goes swimmingly. So why does that stick out? Because it's a it's a request from a customer that seems unreal. It seems it's it's not even it's not even ⁓ understood at the time. But this process of understanding it, tuning, ability to customize, be able to use that real data feedback from the site and say, okay, this is what we're up against. This is what needs to be done. Some of our application engineers can feed that back to the the the training method and we can deploy a new system overnight. ⁓ is just amazing. The the story sticks out in my head because it's just a a both an amazing experience to be part of that, to see it, to see success, but just to see the how far we can push the technology. And and and that leads into that is the process. That is you you adapt whether it's automatically adapts to the scene, to the constraints that are there, or a human in the loop understanding of the requirement shift and feeding that back to the training engine for a second iteration. ⁓ that's kind of where I see that that that shift happening. You just you say sub millimeter and you think about how small that is and then these big pieces of metal we're able to have that level of accuracy. It's still no matter how many times I see a manufacturing floor, those those tolerances always blow my mind. It's incredible. Absolutely. Fantastic. ⁓ hey Jamie, my next question's twofold. If you can make a bold production prediction about where AI vision manufacturing will be in three years, what would that be? And then secondly, is there a problem or applic application specific challenge today that most people are adamant that it cannot be solved with vision because it's just so complex. But you really believe that with advances in the vision technology that it can be solved in the coming years. What would that be application be? Yeah, interesting. I think you know, over the next three years, you know, what what we're what we're seeing already or and what we're going to be seeing more of is this use of natural language. we see LLMs, ⁓ you know, claw, chat GPT just, you know, in every aspect of our life these days. ⁓ and that's going to be part of the setup, part of the Commissioning, part of the design, ⁓ part of the telling a robot or programming a robot how to move is going to be okay. I don't have to push my thumbs into a teach pendant of a robot anymore. I'm going to describe the vision challenge, and that's going to be incorporated into more and more software over the years. and what does that what does that mean for our deployments? That means that that integration time. starts reducing. That means that the the the the simulation, the the ability to understand how this robot is going to function in a cell is going to ⁓ reduce because all of these things happen with but much less friction. ⁓ on on top of that, you know the other term you're hearing more and more of is physical AI. Okay, what does that mean? That is just combining these foundation models, these these LLMs, these this natural ⁓ progression of technology improvements to AI and machine learning to ac have actual physical use cases, actual connect with robots and connect with vision. ⁓ so that's not so much of an a actual prediction of what's happening, but it's just a a wave of investment, advancement that you'll see more and more capabilities come out of robotics, robotics with vision and grasping. ⁓ It sorry, what was the s ⁓ An application today, a real life application today that we see that a lot of the experts in the factory floor, meeting manufacturing leaders, believe that it's so complex that Vision probably will not be able to solve it. This today applications that require dexterous grasping is something that is very challenging. Something that is, you know, high high detail, high precision, high accuracy grasping and gripping of a part. ⁓ there's a lot of development in in human like manipulators, human like ⁓ graspers. ⁓ And that is something that I see a lot of investment in, a lot of momentum towards. But vision related is there's a lot of challenges with ⁓ highly reflective parts, parts that are barely visible, lighting conditions that continue to change throughout the throughout the the the process. ⁓ but furthermore is is high mix, lots of different types of parts that need to be maybe in one bin or cycling through a production flow. and that ability to for these high mix parts, that ability to quickly set up, commission, and deploy a system to be working with these this new part that's just been introduced, ⁓ that's going to be very ⁓ prevalent. That's going to be ⁓ needed more. And With the advancements of AI, the advancements of LLMs, the ability to explain the vision challenge and provide a reference and to quickly turn over a vision solution that will become readily available. Well, Jamie, ⁓ I think those are the questions we had for you today. ⁓ this was really exciting. I enjoyed learning ⁓ more about a para, even though I work here. I feel like I continue to learn something new every day. ⁓ and really thank you very much for being our inaugural guest on the podcast. Perhaps we'll have you back. Thank you very much. Thank you very much. It's happen I'm happy to be part of this. Awesome. Well thank you for tuning in and ⁓ please make sure to ⁓ subscribe and we'll be back ⁓ with our next episode soon. Thank you. Thanks, Peter.