Dax Garner (3:06)

Yeah, definitely. In fact, I'll kind of start a little bit about why I even joined Cognitive Space. Exactly. So I feel like I'm a very mission driven person. I became an aerospace engineer because I want to go into space. I went to NASA JSC because I want to put humans into space. But after working there for a while, I realized that there's an entire infrastructure that needs to go into space that will allow all of us to go into space one day and kind of combine that with watching the AI ML space grow and advance. It occurred to me that space is hard and AI and ML technologies can really make it easier. And that's a key component in getting infrastructure and eventually humans into space. So I joined Cognitive Space because that was essentially the mission and it's not human spaceflight, but it is the technologies that are going to allow us to manage all of the assets that we need on ground and in space that will allow humans to fly if they want to in the future. So that was one of my primary motivations for joining Cognitive Space. Cognitive Space's mission is to empower the use of space assets, particularly developing AI and ML algorithms for proliferated systems and mission management of those assets. So when I talk about proliferated systems, that's hundreds of satellites all working together to achieve a mission that could be taking pictures of the Earth. It could be establishing a mesh network around, a global mesh network around the world. But they've got to work together in order to achieve those missions. And the that becomes a huge optimization and combinatorial space. And that's where our algorithms come in.

Maria Varmazes (5:02)

Tell me more about that. That is a heavy lift. But at the same time, the technology that's available now, I imagine is just massively enabling that. And I'd love to hear more about what that looks like.

Dax Garner (5:13)

Right. So when you're solving combinatorial optimization problems, the idea there is that. I guess, Let me use an example.

Maria Varmazes (5:21)

Sure.

Dax Garner (5:22)

So if you have 100 satellites in the sky and their job is to take pictures of the Earth every single day, you want to make sure that you take the best picture for each one of those targets. But you have multiple satellites that could fly over that target at any time. And so you have a choice which satellite is going to take that picture. So it becomes optimization problem. But when you have a huge combinatorial space, so many options, it can become very difficult to optimize that effectively with traditional constraint based or other operations research type algorithms. It's NPR problem. It takes a long time to solve, if it can be solved at all. Conversely, and historically because of that problem, people tend to use heuristics which are just simple, you Know the first satellite to go over that target takes the picture. It's a simple algorithm, it just gets the job done. But you lose a lot of optimality with these simple rules. And the sweet spot is training and designing ML models that can run at the speeds of heuristics. Heuristics run really quickly, but you can buy back a lot of that optimality, a lot of that performance in your mission and get a lot more pictures.

Ed Miletian (6:47)

There's another layer of complexity there, right when your schedule is dynamic. What if you're losing tasks or gaining new tasks, or what if the thing that you thought was going to take the picture actually can't because of some hardware failure? Now you have to redo all the planning that you've done and re optimize on the system, otherwise you're dropping collections.

Dax Garner (7:07)

Yes, yes, exactly. Another good example is ground communication planning. They plan like long two week cycles and that becomes their reference schedule. But the plan you made two weeks out isn't going to take into account that that antenna has decided not to work today and now you have to replan. But that optimization algorithm that you use to generate a 2 week schedule takes 3, 5 hours to generate and you don't have that time to replan. And that's where technologies like ML can come in. You can reoptimize very quickly.

Maria Varmazes (7:46)

Yeah, the word speed and scale has come up a lot lately. And that's what everybody wants, is what we're moving towards. But then it becomes we have that added complexity and how do we enable that speed and scale without having to rely on these algorithms to take hours, which we don't have. So what you're talking about, I imagine would also enable a lot of really crucial missions that are going on right now as well as future. Can you tell me a bit about that?

Dax Garner (8:07)

Sure. So Cogni Space is working primarily with the Space Development Agency sda. They are building their Mesh network constellation and we are helping them optimize their link management. So understanding which satellite will communicate with another satellite and it's demand based, which is the extra component that's really where ML can help with and is understanding where you want to serve communications demand around the world and how that informs your link schedules. If nodes go down, you can then replan them and still service the demand. So that's one of our primary customers. Another one is NGA and NRO in terms of understanding geospatial requests and servicing those with commercial providers. So they have their own assets that they do mission planning with, but they want to complement those capabilities with what commercial providers are doing. Planet Umbra, isi, Capella, et cetera, Airbus. And we can help them understand that capacity and make predictions about whether they can fulfill certain requests on the commercial side.

Ed Miletian (9:24)

Yeah. I personally of want to highlight what you said there with the geospatial insights. A lot of these folks have very tight latency requirements where they have to shorten the time from a collection down to a dissemination. And sometimes these plans are great when you have all the antennas that you want. And so, you know, if I take an image here, I'll have a downlink opportunity in 10 minutes and I'll get all my data down. It's not always the case because if that antenna, like you said, is gone, it would have actually been better to have a satellite that's lagging do the collection because now the original satellite has to go all the way around the earth to the next contact, which could increase your latency by an hour or two hours.

Maria Varmazes (10:09)

Right.

Ed Miletian (10:10)

So that's why the optimality and also the reacting to new stimulus is really, really important.

Maria Varmazes (10:17)

So I'm gonna go back to the speed and scale. I'm so curious how AWS technology comes in here to enable all these incredible missions that you all are doing. Because I'm just thinking about the heavy lift involved to make all this happen and I'd love to hear a little bit more about that.

Dax Garner (10:33)

Sure. So cognitive space uses EKs and ECSs in terms of running all of our ML algorithms and standing up our applications in their cloud environments.

Ed Miletian (10:46)

Yeah. And I can add a little more to that.

Maria Varmazes (10:48)

Sure.

Ed Miletian (10:49)

Like in terms of the architecture that they've chosen, it scales really well to train the models and also to execute the models when you're planning. And like you said, scale is really important, both in the sense of just the total amount of compute I have, but also being able to onboard new constellations, new missions, and not having to re architect the whole system. That's why deployment like this is really, really crucial for these critical missions that our governments have. Because as they want to integrate constellations, they don't want to go back to the drawing board and have to figure out all these different interfaces that now have to be made. They just want it to work.

Maria Varmazes (11:30)

Yeah, yeah. And I would imagine security is also extremely important given the customers that you've mentioned. Having that baked in is just a given.

Dax Garner (11:38)

Yeah, absolutely.

Ed Miletian (11:40)

Yeah. These services all can run in both our GovCloud regions as well as our most classified secret and top secret regions.

Maria Varmazes (11:51)

Dax, I would love to know about Real world impact, if we have any examples, I mean, obviously not for the national security customers, but just in terms of like, do we have any numbers about efficiencies, like anything like that? Any data set?

Dax Garner (12:04)

Yeah. So as we benchmark our ML algorithms, we benchmark them against heuristics in terms of understanding solve time performance, and then we also benchmark them against traditional operations research, constraint programming type solutions. So that way we understand how much performance we are gaining back in terms of. And so depending on the domain, whether it's network management, sensor planning, we tend to see that our approaches can run a little bit slower than heuristics because it is still an ML model that's running, but we can gain back about where heuristics might perform at 50, 60% of optimality. We're really getting back to like 90, 95% of the optimum solutions, depending on the objective and the constraints. What that means for operational real time performance is that we're planning in minutes. You're planning hundreds of satellites in minutes, whether that's establishing link schedules or planning sensor management and collecting pictures on the ground.

Ed Miletian (13:22)

Yeah. And doing that very quickly is important if you want to keep track of a lot of different areas of interest. Right. So if you're interested in imaging the entire earth with hundreds of satellites, this is a big problem. And so being able to run it very closely to a heuristic model is really important. And on top of that, 90 to 95%, that's really impressive.

Maria Varmazes (13:48)

Yeah. I'm curious, where do you see, with AI developing all the time, where do you see things going from here?

Dax Garner (13:57)

Yeah. So we focused a lot on these specialized ML algorithms that solve big combinatorial space optimization problems. But fundamentally I think that those are just tools for agentic systems. What is your constraint today might be your objective tomorrow? And when we talk to operators, they're planning reference schedules to just make sure that they're going to meet their operational needs. But then as dynamic things come in, they might need to just get. I guess a good example is data latency, like you had mentioned, can be paramount. And so you want to minimize your general data latency. But as resources get constrained because a network node goes down or whatnot, you just want to minimize your latency. So what was just a constraint of fundamental operations? Constraint becomes an objective function. And why that matters is because we will train machine learning algorithms on an objective function, on different objective functions and different constraints. And when you combine those models, those optimization models, and pair them with an agentic system, an agent Gets to the idea is that an agent will start to decide. I'm going to use this model today because my operator is asking for this and I recognize that a node is down. So this model might be the best way to resolve and replan. Maybe I have time to generate a reference schedule so I can take the time to run a constraint programming solution. Maybe I don't have any time. And the greedy heuristic is the best to get a plan out right now because I don't care about maximizing performance. And so it's providing agents these tools.

Ed Miletian (15:54)

Yeah, the agentic piece is really interesting and one thing cloud is really good at is it integrating these systems together in a common platform. So like now that you have this model, it's much easier to build these agentic workflows on top of that. And that's something that we had really focused on building out our cloud Mission Operations center concept, which is mission operations in the cloud, like the name suggests. So what that is is being able to run the various subsystems of a mission operations system at scale. And so that's flight dynamics, mission planning, command and control, data processing, as well as orchestration. The benefit of the cloud Mission Operations center is you can run best of breed solutions at scale, like cognitive solution. So if you have a big problem, you can scale up your mission planning. You don't need to scale up everything so simultaneously. And you don't need to worry about how big will my system have to be over the next 20 years. You're just solving the problem that you have today and you know that it'll grow to meet your needs. The other benefit of running your mission operations on the cloud is that you're always getting access to the best underlying infrastructure. You don't have to worry about provisioning new GPUs, new CPUs, and you don't have to worry about getting rid of your old stuff, which in classified systems is a big problem because that stuff has classified data on it. You can't just wheel it out and throw it in the trash bin. There's a long process to get rid of it and procure new hardware. With the cloud Mission Operations center, you are automatically getting the best available technology under the hood, both at the infrastructure level and at the application layer.

Maria Varmazes (17:42)

Yeah, anything you want to add to accident?

Dax Garner (17:44)

I guess I'll just say that we definitely see aws. AWS as key to helping us move our technology from unclassed to classified. Supporting common cloud native infrastructure has been key.

Maria Varmazes (17:57)

I appreciate that. I know that we're Coming up on time, I want to make sure that I give you both an opportunity to have the floor have a wrap up. Is there anything you wanted to add Dax, about what Cognitive Space is doing or what you're looking at for future missions?

Ed Miletian (18:12)

Yeah, I think just looking forward Constellations are going to keep increasing. That is just a trend. That is absolutely true. So do customers want to continue to worry about buying more and more infrastructure, doing trades on, you know, do we bring this stuff in? How do we grow? AWS is gonna be focused on reducing that burden on customers so they can focus on mission and as well as providing a platform for our partners like Cognitive to build out to meet the customer where they're at and provide those critical mission services.

Dax Garner (18:50)

Dax yeah, I think I will add that we're very focused on the US Government at the moment. They're building the most proliferated systems and I think these technologies will be developed there. But the key is to enable an entire space economy. And as startups, there's many companies out there that want to also fly constellations of satellites to do really cool missions. But they're startups, they're focused on putting their first spacecraft in space. They're not thinking about how to manage a proliferated constellation. And that's where a system like ours that can do that, that's already designed for proliferated system and can help them mission management. AWS's Cloud Mission Operations center is already there ready to onboard their proliferated systems as they fly their first, second and then eventually become their full scale constellations. That's really the future I think in our collaboration for sure.

Maria Varmazes (19:52)

Well gentlemen, thank you both. It's been a pleasure.

Dax Garner (19:54)

Thank you. Thank you very much.

Ed Miletian (19:54)

Thank you. Foreign.

Maria Varmazes (20:03)

That'S it for this episode of AWS in Orbit by N2K Space. We'd love to know what you think of this podcast. You can email us@spacen2k.com or submit the survey in the show notes. Your feedback ensures that we deliver the information that keeps you a step ahead in the rapidly changing space industry. This episode was produced by Laura Barber for AWS Aerospace and satellite and by N2K producer Liz Stokes and senior producer Alice Carruth. Mixing by Elliot Peltzman and Trey Hester with original music and sound design by Elliot Peltzman. Our executive producer is Jennifer Ibin. Our publisher is Peter Kilpe and I've been your host, Maria Varmazes. Thank you for joining us.