209 - AI, Spacecraft Autonomy and Space Data for Humanity

Podcast from Kratos. My name is John Gilroy and I'll be your moderator. Joining us today is You'd Architecture, aerospace and satellite at AWS. AWS, a cloud computing platform by Amazon, launched nearly two decades ago, provides services for organizations. that need to store and process sensitive data. Yudhajit is here to share his knowledge on cloud computing for space and all the ways generative AI can help make life on Earth more efficient for humanity. Ready to jump in here, Yudhajit?

Sounds good. It's an honor and a privilege to be here. No, I don't know about that, but it's gonna be fun, we know that. So when looking at the space industry And how much it's changed in recent years, how has the technology of cloud computing contributed to that growth? Oh yeah. What a what a great question to start off with. Um I would say cloud has been one of the most powerful accelerants in the modern space industry.

It has helped change our perspective uh to a mission first model from an infrastructure first model. And just to expand upon that a little bit, um on the mission design side, now we're seeing customers Now build and test space systems over cloud based digital twins. and other collaborative simulations. They are also using high performance compute systems to run orbital dynamics at scale without a need for on prem servers.

Uh and then going on to mission ops, we have cloud native pipelines that allow ground control, uh contact scheduling, and even anomaly responses to be automated and accessible from anywhere. And then then there's the data part. And John, that is where cloud really shines.

And I love talking about our customers and I'll I'll try to do that, slip in a customer reference here and there throughout our conversation today. Uh the one I wanna talk about is Capella Space. Earth observation companies like Capella Space, they went from a six hour SAR delivery window to under thirty minutes.

Using AWS and I'm not just talking about um GPU powered data processing, but in fact every stage of the chain, like like serverless tasking, like global content delivery networks, and of course a robust security infrastructure. They all allow raw satellite data to be converted into actionable insights within minutes.

And because I'm in a technical role I always start talking about technical stuff. Yeah. Uh but it's not just the technology that has grown grown so much. The the business model has evolved as well in the industry. Um because of the the cloud's pay as you go, nature, startups no longer need to invest in uh in antennas or in data centers. They're operating entirely from the cloud and that lowers the barrier to entry.

and just opens up new opportunities for innovation all across their value proposition system. New opportunities for innovation. I like hearing about that. I would think that most of our listeners know the Amazon story, the origin story. A guy trying to sell books and all of a sudden he got lucky and successful and branched off into AWS. We know that. But tell me about the origin story.

behind ground station in the cloud and w why and and how did it even come to be? As the fundamentals of Amazon, we work backwards from our customers and this was no different. In fact the idea of ground station in the cloud it came from listening to a repeated frustration from our satellite customers that the ground segment was slowing them down.

Now traditionally you will know very well that satellite operators had to either build or lease antenna networks, which would mean along lead times and fixed locations, a high capex. And from a cloud perspective, a really a fragmented integration with their cloud environments. And they would say things like All of our data gets delivered to our S tree buckets. Wish there was a way to capture it closer to the antenna locations. Oh that disconnect, John, is what led us to ask.

Can we treat the antenna like a cloud resource? So in twenty eighteen we launched AWS Ground Station, which is a fully managed service where you schedule satellite passes. through API, receive data directly into S three and trigger post processing workflows within milliseconds. And one great example customer story like I like to talk about is uh DORBIT who are using ground station as a service to orchestrate downlinks for multiple payload customers.

What it's done is it's turned a hardware bottleneck into a software service. And the moment it's a software service, we can then scale it globally using DevOps principles and and cloud fundamentals.

You said you earlier you said that you love data, so I'm gonna have to ask you a data question here. So you know, with the cloud comes data, and with data comes a resource for predictive analytics, machine learning, and generative AI. So what kinds of data are being stored and processed? in support of these technology enablers? Of course. Great question. Data has gravity and uh I like to split data, just for my own sake of understanding, into three categories.

Uh one is what's happening on the spacecraft, data from there. Second is the observations on Earth or in space that we are making using the sensors. And third is how those systems are behaving with each other over time. So really quickly, on the spacecraft data we are capturing what? We are capturing telemetry data, right? Voltage curves, power cycles, A D C S data and all of these are a rich time series data sets that are fueling these predictive maintenance models.

Uh from the preload side we get uh you know the standard uh imagery data, RF data and science data. And so often they are in unstructured formats which then requires intensive processing. And that's not not to mention that we also have layers and layers of operational data like command logs, like scheduling patterns, and that is exactly where machine learning comes into play. Because time series models they detect the delta from the usual in subsystems much before human operators would notice.

And it's not just analysing the data, we are training our systems on it. So much so that the satellite is not just a sensor anymore, it's very much a part of a feedback loop. And this is how as an industry we have progressed towards generative AI. You're training the ML models, they're ingesting the data, and now generative AI can summarize and contextualize and even suggest actions and next steps.

Next week I'll be going to the trade show in downtown DC and D C everyone talks about security. It's a really a hot topic. Number one on the list here. But important for everyone over the world, we know that. So it's gotta be a high priority to keep all this information secure. So can you explain how it stays safe? Absolutely. Security is top priority. Security and compliance are our highest, highest priority. It is fundamental to how we architect our infrastructure and our services.

We approach uh the topic of security through what we call the shared responsibility model. where we handle the security off the cloud and we provide our customers the tools and guidance so that they can secure their assets in the cloud. We uh maintain dozens of compliance certifications and accreditations worldwide, uh like PCI DSS level one.

uh like ISO twenty seven thousand one and FedRamp. And we have we have built security into every layer of our operations, from physical data centers to network architecture to to processes. And we provide native security services and features that integrate seamlessly, seamlessly with our colour with with our customers' workloads. And not just that, we also uh we also support our customers' existing security tools and controls.

And of course the the uh the landscape of security evolves constantly, which is why we maintain an innovation cycle where we are launching hundreds and hundreds of new security features every year and trying to stay ahead of emerging threats.

You Hajip, let's take a look at two topics here. Spacecraft maintenance and or ground infrastructure. So does it work the same for the way it does for cars, you know? Some people even get text messages when they gotta get their oil change in their car. Is that how it works? That's a great analogy. Um we just got a car.

And uh and I'm looking at the app and I'm checking whether it's time for an uh oil change. Uh no, I drove it for a week, so probably too soon. A little bit. But no no, great question. I would say um uh there are overlaps And there are differences. And then let's talk about the overlaps first. AWS has published a guidance for aircraft predictive maintenance, where aircraft sensors, they stream data into the cloud via Amazon kines kinesis.

And then there are ML models in SageMaker that flag any kind of anomalies in real time. And whenever a failure pattern is recognized. the system automatically pushes a targeted notification to a technician with the component ID, with the diagnosis and recommended actions. Now, our customer, LiveEo, has implemented a very similar workflow. What they are doing is they are they're using satellite imagery and machine learning to monitor vegetation risks or storm damages.

What the what the system does is it flags high risk segments and dispatches field crews with GPS coordinated work orders, which is essentially like your maintenance text, just on a on a national scale. Uh that is the overlap and the differences uh the main one has to be scale, right? You build you might build ten thousand cars a day, but only like ten satellites a year.

the volume of data is much smaller than in aviation or uh or automotives, right. So we lean more on on physics based models like digital twins. And now leaning more and more on unsupervised or reinforcemental learning uh to bridge that gap. You know, um you mentioned pattern recognition. Let's take this and take it to outer space. Okay.

One step further, huh? So how can AI help with collision avoidance either with other spacecraft or with space debris? That, John, is a is a topic of frequent discussion. Just just in low Earth orbit we are tracking what about Forty thousand objects and and the number is growing fast. Um traditionally collision risk used to be assessed using deterministic models, right? Built around TLEs and orbit paths. Uh these models often assumed ideal conditions.

And didn't account for uncertainty of debris, for example. And uh there used to be a lot of false alarms as well, which would uh cause unnecessary maneuvers, right? Uh that's where machine learning and artificial intelligence just changes the process from from reactive to predictive. And another of our customers, Leo Labs, is is a key player in this space. What they are doing is they're using ground based phased array radars to track these debris in real time.

Uh they collect the data and then they feed into the machine learning models and that calculates probabilistic collision risks and that goes, you know, well beyond just simple proximity alerts, right? And just these these models and these companies are helping us just not to dodge space debris. Yeah, that's there. But just truly laying the groundwork for a more safe, a more autonomous traffic management in orbit.

Let's go from space debris and go down to the ground. So what are some examples of the other? So one of the most relevant I think and the technologically brilliant examples is is from our uh customer ISI in Europe. who are redefining how we respond to floods and other disasters. So what ISI is doing is they are they produce uh detailed flood extent maps using SAR imagery. And then they make this data actionable through a Gen AI assistant uh using Amazon bedrock.

And this assistant lets users ask natural language questions and get actionable responses from it. And to power this system, ISI uses a host of other AWS services, like Amazon S three to store their map. Amazon uh SageMaker to that builds and fine tunes the geospatial analytics portion. And what else? They use Amazon Athena, uh, which allows the generative AI assistant to query structured flood impact data on demand.

Uh and of course Amazon Cloudfront and API gateway that deliver these insights quickly and securely. to the disaster response teams. So thank you, John, for for this particular question. My wife works in the renewable renewable energy industry. Yeah. And we are passionate about uh making the earth a better place to live in. Uh what ISI is a fantastic example of of how they're doing that and helping people who need it the most.

Well I heard it's summertime, people are going traveling on vacation. In August I'll be visiting a farm in the state of Washington. So let's talk about farming and agriculture, you know. Let's talk about this, you know. That sounds fun. Anyway, so what about agriculture and these farmers, you know? Can you share some examples? of how AI is benefiting the ag industry and and humanity writ large. Absolutely, John. Ours is a global team in AWS.

And uh w we have the privilege of working with customers all around the world, right? And specifically in the agriculture space, I wanna highlight the work of a company called Degas, which is based out of Japan and they are doing some great work in Ghana. Degas is supporting farmers in Africa, many of whom are are vulnerable to climate changes and lack access to financing and agronomic knowledge. I'll give a quick synopsis of how this works.

Degas collects two streams of data ground level information through their in house Android app and satellite imagery, both both optical and radar. And the data is then pushed into the cloud on a data lake in AWS, where there are models to do a number of things, like calculate field level crop health indices and predict crop yields. But John, the the real innovation?

is how they're using generative AI. Degas has built a chatbot using Amazon Bedrock along with services like Amazon Kendra, Athena, RDS, and this chatbot allows agents in the field Uh to ask natural language questions like should we apply fertilizer this week based on rainfall and vegetation health?

Uh what excites me so much is that it's it's not a data pull from static documents. The chatbot is generating responses about That specific plot of land, it's so customized, that specific agronomic protocol in that area and then giving expert level guidance in the palm of the farmer's hand. And the results are amazing, they're tangible. Over ninety percent of their farmers.

Saw income increase in the And this is such a perfect example of how AI is expanding opportunity, is expanding equity in some of the most climate vulnerable regions on Earth.

Well we're gonna yo yo again. We're gonna go from tractors in Africa to outer space again. So let's talk about AI again and go to outer space. So so how does AI support You know, the future journey, man's journey into space, to Mars and beyond. Infinity of beyond.

Uh human space exploration always uh such a fascinating topic. Uh we have we have talked about l n large language models, right? Uh I gave the examples on large language models so far, but for this one I want to briefly mention small language models or SLMs that uh are running directly on spacecraft or rovers. Because in my opinion, John, that is the biggest constraint for deep space exploration. It's not propulsion, but it's spacecraft autonomy and it's the latency of communication.

Now in the telco industry this is already happening. We are already deploying those SLMs on device. with something called an agentic orchestration for hyper personalized customer experiences. And the same pattern applies to space as well. So instead of routing every anomaly back to Earth, What a satellite does is it will analyze telemetry on board and generate resolutions. We are working on these architectures.

Uh which are a combination of retrieval augmented generation or rag, local telemetry, and at the end of the day it's outcome-based planning. What this does is it it sets up the foundation for agentic autonomy in space. Where AI models, based on their data, based on their goals, and based on the environment they are in, they do the orchestration of action. You know, you should five years ago

I would never have th put these two words together. Agentic orchestration. It just wouldn't it wouldn't have made any sense at all. Now people just say, Oh yeah, it's agenic orchestration. So five years from now, looking into the future, so what are some ways You think generative AI will help us that it isn't already? Ah that's uh that's that's a tough one. Sean, because um because we work with these technologies every day.

Uh we see its applicability everywhere. But I'll I'll give you a personal opinion, right? Um I think today what generative AI is, is it is an interface between humans and systems. And I I believe that the next stage of innovation would be when generative AI becomes an interface between multiple systems working with each other. Just like just like a constellation of satellites making decisions between themselves.

You know, sharing context, negotiating priorities, generating action plans and even even resolving conflicts. And that level of collaboration, it requires Gen AI models to understand intent, uh understand the missions. And most importantly, John, uh they have to understand the consequences of their actions. But I believe uh we are not too far away from that day. You did g I think you have given our listeners a real good understanding of artifacts.

I'd like to thank our guest, Yusajita Skupta, Head of Solutions.