Episode Transcript
[00:00:00] Speaker A: Companies like ChatGPT have used new types of AI models that can digest all kinds of inputs, not just 2D images like our AI models, but they can take all kinds of information from the sensors that the robot has. They can take that information and they can understand it, and they can begin to tell a robot what to do next.
[00:00:30] Speaker B: Hello everyone, and welcome to the Robot Industry Podcast. We're glad you're here and thank you for subscribing. Before we get to our episode today, I wanted to let you know that I have a new podcast I'm starting called Automation Matters. It's about the front end of the automation business. And whether you're a builder, integrator, distributor, or robot oem, this is for you. This new podcast is about sales, marketing, business development, strategy, and much, much more. And I'm excited about Automation Matters, but more on that later. Our podcast for today is about hype versus reality, an important approach to foundation models like delivery robots. And I'm thrilled to have Sean Edwards on the podcast today. Sean is the Chief technical officer at PlusOne Robotics. Hey, welcome to the podcast, Sean.
[00:01:12] Speaker A: Yeah, I'm really excited to be here. I'm really excited to talk about advanced AI in robotics. And at least from my view, I see as what's realistic and what's maybe a little bit of hype.
[00:01:26] Speaker B: First, let's get started. For people who don't know your background, can you tell us a little bit about your experience at Southwest Research Institute, which also ended up to be Willow Garage?
[00:01:37] Speaker A: Sure, sure. So I have about 20 years of experience in robotics, much of that at Southwest Research Institutes, and I started my career there.
It's an amazing place where we do at Southwest Research, we did research for the government, we did research for commercial companies, and really if you had a hard problem to solve in robotics and automation, you would go to Southwest Research Institute. And I had the opportunity to work on the largest robots in the world. The robots that strip paint off of fighter aircraft, truly advanced computer vision applications, state of the art for the day. And it was just a great place where at all times you could look for the state of the art or the bleeding edge of technology and figure out how it applied in the real world.
One of those times was when ros, an open source robot operating system, came into being. ROS came out of Willow Garage, which was a Silicon Valley based startup, and their goal was to develop a really core advanced platform for robotics development. With the work I was doing at Southwest Research Institute, I realized that that was what was missing in terms of Industrial applications. There were certainly industrial libraries for robot control and computer vision, but they weren't that capable and they weren't at the scale that I felt they could be. But the unique thing about ROS being an open source software package is that it could be at that scale. And certainly that's been borne out over time with all the different companies that have utilized ROS for all types of robots. But that was the beginning of it. And the cooperation between Southwest Research and Willow Garage to develop Ross Industrial, which, which really just brought ROS to industrial applications like the ones you and I get excited about.
[00:03:47] Speaker B: And for some of our audience who don't really understand the importance of ros, like ROS is one of these tools that lots of startup robots use, right?
[00:03:57] Speaker A: Yeah, yeah. I mean it's in any product development or any kind of research or advanced technology development.
One of the ways you're limited is by your tools. And I'm not sure everybody fully appreciates that. If you're an engineer and you're down in the trenches, you probably do, you know, that your tools are really important.
I always felt when I was at Southwest Research Institute that we were limited by the tools that industry gave us. Every robot had a different programming language, the computer vision libraries. They existed, but they were expensive and somewhat limiting. And Ross kind of solve those problems. It was open source, it was freely available. It brought in the state of the art technologies across path planning, computer vision, kinematics, AI. All of those things would become freely available and at a scale that could support it. With this powerful tool set, you could now do really advanced things. And that's what I think has kind of spawned really advanced robotics across warehouse agriculture, you know, in the factory. Just lots of things that, you know, over the past decade have, have been enabled by Ross.
[00:05:14] Speaker B: And there's a lot of industrial robotic companies who are in on this too. Right. Like this isn't just you and, and a few other companies. This is a big thing.
[00:05:23] Speaker A: Yeah. So I mean it was 12 years ago now that I was at, at at Willow Garage developing the core libraries that would become ROS Industrial.
As part of that we engaged industrial companies, industrial robot companies, big manufacturers. And we said, look at this cool software, look at what we can do with it. And it's really going to advance robotic automation.
That group of companies became the Ross Industrial Consortium and it really helped push and guide the uses of ROS in kind of our robotics manufacturing type applications.
[00:06:05] Speaker B: And now I'd like to just. So now you don't work for will garage or ROS anymore. You actually work for a company called PlusOne Robotics. And if you could, I think it's important to the conversation, if you could just tell us a little bit about PlusOne Robotics.
[00:06:18] Speaker A: Yeah. So PlusOne Robotics is a San Antonio based startup.
We're about seven years old, eight years old if you count the time when we didn't get paid. We started the company with the idea that Ross as this enabling technology would bring advanced vision and intelligence to warehouse robotics. We were wanting to be really focused on warehouse robotics.
That is what, what PlusOne Robotics does. We do things like parcel sortation, manipulating packages. We break down pallets, we palletize items onto pallets. If you're in a warehouse, anytime a person is picking up a case or a package, that is, that is a target for our technology and our robots.
[00:07:11] Speaker B: Vayu Robotics recently announced its ditching lidar. And LIDAR stands for light detection and ranging technology and leveraging foundational models for their newest delivery robot. So maybe you could tell us a little bit more about LiDAR and then talk a little bit about what we're going to talk about today.
[00:07:29] Speaker A: Yeah. So LiDAR is just one of the sensors that kind of give robots a view of the world. There's certainly 2D cameras, there's 3D cameras that give you point clouds. So it's all this technology that essentially gives the robots the ability to see. I think when you talk about robots, most people just assume that they can sense and perceive their environments. That's not necessarily true of robots in factories. A lot of them are completely blind. But these new sensors really give them that ability that most people would assume they have but didn't.
[00:08:08] Speaker B: So this announcement comes on the heels of of Covariant, another company, Covariant announcing something called RFM1. It's Robotics foundation model that promised general intelligence for robotic control access across tasks and environments earlier this year. So what's the challenge there and what's a foundation model? Maybe.
[00:08:30] Speaker A: Yeah, let's start with a foundation model. So PlusOne Robotics, one of the core technologies that we develop is a computer vision system. The core technology behind that, AI models and convolutional neural nets. But essentially AI focused on the task of taking that perception data, mostly 2D images, but using that information and identifying objects within the scene. That's a core technology.
That's how Plus1 utilizes AI. But that's kind of old tech, right? That tech came along like five years ago, but it was a tremendous leap in computer vision technologies. What has happened since then is that companies like ChatGPT and others have used new types of AI models that can digest all kinds of inputs, not just 2D images like our AI models, but they can take all kinds of information from, from the sensors that the robot has. They can take that information and they can understand it, and they can begin to tell a robot what to do next.
And that's when people are talking about robotics, foundation models. It's less so about.
It's not like a ChatGPT model, but it's this idea that an AI model that can ingest all this information and tell the robot to intelligently do something about it. Today, that's brute force by programming and software engineering, and it's somewhat limiting. Again, us roboticists are limited by our tools, and foundation models are going to be a new tool that I'm really excited about because it's going to enable new capabilities, new use cases, because it is a step change in the intelligence of the robots that we can deploy.
[00:10:31] Speaker B: So what's it take to create a foundation model? What do I need?
[00:10:34] Speaker A: There's kind of two parts to it. So first, you need an architecture, right? I would say that when we talk about ChatGPT and what that kind of meant for the work that we all do, right?
It takes in text and it can give you an answer. Part of what enabled that was having an architecture purposely built from that for a foundation model. We need that architecture. Now, some people are trying to leverage transformer models, which ChatGPT is based upon as one type of model architecture, but there might be others. But whatever it is, that model architecture is kind of the core, the structure of how the AI will reason and make decisions. So you have to have that. And then the second thing you have to have is lots of Data. That's how ChatGPT had access to all the data on the Internet. And that's kind of how it trained, it became generalized, and it could give us answers to questions we had never asked it. Basically, well, the same has to be true for robotics, or we believe the same is true for robotics, which is if given enough information, enough data, and the right model architecture, we can create truly a brain for a robot, such that when you put it out in the field and it's sensing the warehouse around, knows what to do, it knows what actions it can take, it knows the job I'm in is parcel sortation. So I got to pick up a package out of this pile of packages and I place it on the conveyor one by one. Whereas today we hard code or program all of that the promise of foundation models is that knowledge will be encapsulated in the AI because of the model architecture and the data that, that you've taught it with.
[00:12:23] Speaker B: And so we're just in the really beginning of foundation models. Is that true?
[00:12:28] Speaker A: Yeah, yeah, the very, very kind of, kind of upfront piece of it. Everybody's excited because they saw what ChatGPT, how quickly ChatGPT moved and how, how, you know, what new kind of functions and capabilities that came about quickly when people kind of saw that. And now people are saying, hey, if we apply it to robots, what new things will our robots be able to do? But I do think we're getting a little ahead of ourselves. Don't get me wrong, I'm really excited about these foundation models. I think it's going to be a step change in what our robots can do. The part that I think is difficult is to figure out when that's going to happen. And there's a really big difference between ChatGPT and, and any kind of foundation models. And it comes down to data. The data does not exist at the scale and the level for robotics that it did for ChatGPT.
That is something that is clearly going to hold back or at least delay the true power of robotics foundation models.
[00:13:45] Speaker B: So some companies are starting to use this technology already.
[00:13:49] Speaker A: Yeah. So as you pointed out, covariance is 1, covariant was recently acquired or that technology was acquired by Amazon. We know that's public, so we know they're interested in that. Ambi, just as recently as like two weeks ago, who's another robotics company in the warehouse space, has also announced some foundation model work. And at Plus One Robotics, we have over 1.4 billion billion picks that our robots have done. We have made it a point to capture a lot of that data. And so we too are interested in foundation models. And we kind of believe our unfair competitive advantage is that we have a lot of data that others might not have.
And so there's a lot of companies pushing towards it. I believe the ones that have the data are going to be the ones that make the most progress, but we're certainly not alone in that.
[00:14:51] Speaker B: And so what's the sometimes challenge of maybe using a foundational model without enough data?
[00:14:58] Speaker A: Yeah, if you don't have enough data, there's accuracy issues.
When the robot senses something, it does the wrong thing. So. So that can happen. And that happens with ChatGPT, by the way. Sometimes it gives you the wrong answer. But the problem is in robotics, the robot doing the wrong thing can be dangerous, it can damage things, it could damage the robot. So it's, it's pretty risky to kind of go to production with, with a model that we might call half baked, right? So you have to be very, very careful about doing that. And to my knowledge, I don't know that anybody has gone to production and truly been in production where they could depend on the AI to address all concerns.
[00:15:54] Speaker B: So what is the lifespan or lifetime of a foundational model? And do we really know that yet?
[00:16:00] Speaker A: Like I said, there's two parts to it. So the more data you get, you can take the same kind of model architecture and you just make that AI model better, right? So to me, that's not kind of, you're not really iterating, you're just taking one architecture and making it better.
So to me, the lifespan of a foundation model really is when that model architecture evolves. And you can kind of think of just like different levels of intelligence. Right? So, you know, being able to respond to questions yes or no is one level of intelligence, being able to kind of provide factual data to support whatever conclusion you're drawing. And all the way up to, you know, a logical proof those, those are different kind of types of, you know, levels and scale of models. And the same will happen in robotics, I bet you. So for example, I do believe that the first foundation models will be very hyper focused. They won't be able to do every job in the warehouse, but maybe one does parcel sortation and a different one does depalletizing or palletizing operations. The real excitement, by the way, will be when you get one model that can do them all.
That's truly going to be exciting. But I don't think that's going to happen very soon and certainly takes a lot of data to get there.
[00:17:23] Speaker B: I have a question about where are we in the cycle of foundational models? But I guess you'd argue we're in the very beginning.
[00:17:29] Speaker A: We are. You know, we started this conversation, we said, you know, what's hype versus not? I think if you're going to talk about foundation models, you know, I'm excited about the work that plus one does in the warehouse with more traditional robots. But I think you also have to talk about humanoid robots. Right? There's a lot of excitement around humanoid robots and when you really dig in, you ask them like, what's the promise of a humanoid robot? Because isn't a purpose built robot better for certain tasks? While that may be true, the excitement around a humanoid robot is that a lot of Tasks in this world are set up for people.
People can get to the workspaces, they can do the manipulation tasks that they need to do, and then when that job's done, they can quickly move on to another job.
When you think about that, what has to empower that? Well, it has to be that generally intelligent robot. That thing that I said, hey, maybe that's further away. Right. It has to be true. They have to achieve that. Otherwise a humanoid is a more expensive, less performant robot for doing tasks that other purpose built robots do much better. Right. When we talk about the hype cycle, I really try to remind everybody there's a lot of hype around humanoids and foundation models have to deliver in order for humanoid robot companies to fulfill the promises they've made.
[00:19:00] Speaker B: So I have a question here about what are the gaps? And so I guess we've been talking about the gaps all along, right?
What about the risks and rewards?
[00:19:09] Speaker A: Yeah. So I think the biggest risk associated with all this is you have to collect a lot of data and there are tricks to it. You could simulate a lot of data, you could generate data, there's data augmentation, there's a lot of tricks to creating a lot of data. But the best data comes from the real world and that is expensive data to get.
And so nobody knows how much data we need. Right? We just don't know. But it's probably a lot, which means there's that it's just going to be expensive to get. And then training these models is also very expensive. Right? You can, some of the language models are many, many millions of dollars to train. And arguably a robot foundation model is, is maybe more complex and more millions of dollars to train. So there's a lot of risk in all of that. And at the very end, it may not be any better than our brute force programming that we do today.
That's the risk.
I don't believe that to be true. I do believe that these AI models are going to outperform kind of our traditional engineering approach.
But it's the risk.
[00:20:35] Speaker B: And I guess the big challenge, right, is that we've got lots of people investing in unproven environments.
And they were talking about employees and investors and startups and even CEOs, right. Maybe not understanding some of these big gaps.
[00:20:52] Speaker A: Yeah, I mean, it's exciting to see all of this kind of come in to robotics and humanoids is really driving a lot of it and there are a lot of people that are at risk.
One thing that I think everybody should kind of remember is that eventually people focus on what do robots do.
And as long as these models can do those things, maybe they're not the most general purpose model.
General purpose robot that walks around the warehouse and does every task, maybe it does two or three really well, that will be okay.
But if I were kind of talking to people about who to get excited about, it's the ones that are that quickly figure out what applications they want to do and focus their efforts on there, that mitigates the risk and kind of accelerates your time to market versus a general purpose robot approach.
[00:21:48] Speaker B: So what's next for these foundational models, do you think? Well, obviously people need to be spending more money, more time.
[00:21:56] Speaker A: It's a question of whether current architectures will work as well for robotics. I'm not sure they will. I think we'll have to. We'll come up with new types of architectures that make more sense for robotics versus language models.
That will generally be true. And then it's a matter of figuring out how to collect the data. And because unlike language models that could use publicly available data, this is also a problem we have to solve for robotic foundation models. And it may only be that the biggest and best companies or the companies that are already at scale that can collect that information, it's going to be much harder for those very early startups.
[00:22:43] Speaker B: Do you see a future where companies are changing data, sorry, sharing data so that they become a better foundational model?
[00:22:51] Speaker A: I do, but part of that is I came from an open source background. The open source person to me wants to believe that there could be open data sources that we could all use to train our models. I don't know if that'll happen. To a certain extent it's kind of happened in language models, in that companies like Facebook open source their models. What I think is more likely in the space that we're in in warehouse robotics is companies kind of partnering and there's going to be companies that know how to build the models and there's going to be companies that have the data and they're going to come together and through that partnership really kind of do the early deployments of these models.
[00:23:37] Speaker B: Yeah, I see a whole bunch of companies becoming more and more valuable the more data that they collect, which just makes a lot of sense.
[00:23:43] Speaker A: Yeah. And it's definitely driven values in the past.
[00:23:49] Speaker B: Sean, this has been very interesting. Thanks for joining us. Have we forgotten to talk about anything? Maybe some of the latest development at plus one.
[00:23:57] Speaker A: All of this AI stuff is certainly stuff we continue to push at plus One, we're actively looking into leveraging the data that we have and either developing our own models or finding those partner companies to help us do that as well. But I see this as a parallel track.
While we're doing this, we don't know when these investments in AI and foundation models are going to pay off. We're making them because we know they will pay off. It's just a matter of when. At the same time, Plus One is also, through our tried and true traditional engineering approach, developing advanced capabilities in parcel sortation, palletizing, depalletizing. We just recently released the fastest parcel sortation robot in the world, or at least what I believe is the fastest partial sortation robot in the world.
And we're going into production. It's exciting to see those robots get deployed to solve real problems that companies have today. But also the data that comes back from those it feeds into our AI.
And it's a key part of this flywheel that Plosone is leveraging to truly advance not only our robots today, but the foundation models and industrial robots of tomorrow.
[00:25:31] Speaker B: We've been talking a lot recently about the dull, dirty, dangerous jobs, but I'm adding an extra D to that sentence. It's the, the data.
You've won some awards too, some money. And you've got this new billion picks going right at plus one.
[00:25:47] Speaker A: Yeah, yeah, we've.
Yeah, we've. We've certainly done quite a bit recently. We got 1.4 billion picks. I think the last time we, we announced it, it was just 1 billion, but we're up to 1.4 billion and we're getting recognition for that in terms of being a leader in robotics technologies.
The part that I like, which is customers, customers I can't talk about today, but they're seeing that. They're seeing that, hey, Plus One has this experience and they have this new technology that solves a problem that they have.
[00:26:28] Speaker B: Well, listen, thank you very much for coming on. How can people find out a little bit more? Maybe reach out to you?
[00:26:33] Speaker A: Sure. I think my favorite place to interact with the community is on LinkedIn.
You can just look up Sean Edwards. I'm old enough that I'm one of the originals, but definitely reach out on LinkedIn or certainly come to our website and you can fill out a contact form there. And that'll also make its way to me as well. Yeah.
[00:27:00] Speaker B: And they can't find you. Find me Jim Barrett on LinkedIn. We'll connect you.
[00:27:06] Speaker A: That's right. You're much easier to find. You have a much, much larger network, I'm sure.
[00:27:10] Speaker B: Well, this has been very just been fascinating. Thank you very much for helping educate us and talk a little bit about foundational models. I'd also like to thank our Earhart Automation Systems. They're our sponsor. Earhart builds and commissions turnkey solutions for their worldwide clients. With over 80 years of precision manufacturing, they understand the complex world of robotics, automated manufacturing and project management, delivering world class custom automation on time and on budget. And you can contact one of their sales engineers@earhart automation.com and Earhart's hard to spell. It's E H R H A R D T. And I'd like to acknowledge A three the association for Advancing Automation. They are the leading automation trade association for robotics, vision and imaging, motion control and motors and the industrial artificial intelligence technologies. Visit automate.org to learn more. And if you'd like to get in touch with us at the Robot Industry Podcast again, you can get a hold of me on LinkedIn. And today's podcast was produced by Customer Attraction Industrial Marketing. And I'd like to thank my team, Chris Gray for the music, Jeffrey Brebner for audio production, my business partner Janet, and our sponsor, Earhart Automation Systems.
