NASA & CosmicBrain AI on Humanoid Robots, Physical AI & Simulation | AI Rabbit Hole 2025

Huge thanks to our panelists for sharing all those insights. The next session is a fireside chat on NASA supercomputers and simulating the next generation of robotics. Let's welcome to the stage Anto Patrex, founder and CEO of CosmicBrain AI and Jason Cornelius, aerospace engineer at NASA. My name is Margarita Geleta I'm a researcher affiliated with Berkeley and Stanford and I'll be your moderator today. So, let's start with some introductions. Jason, tell us a bit more about yourself. Sure. Well, good afternoon everyone or good morning. I don't know where we're at. I know it's right before lunch, but I've had enough coffee hopefully for the whole room here. So, I think we'll we'll keep it a little bit energized for you guys. So, I work at NASA as research center down in Mountain View. and my background is in aerospace engineering, working on developing kind of machine learning methods to do aerospace vehicle design optimization.

And I I'll keep it there. I think we'll talk much more about that. Hi guys, my name is Anto Patrex. I'm the CEO and founder of CosmicBrain AI. I used to work at XAI for a while before that. Did a couple of projects at NASA as well. Good to be here. Thank you. So to kick off, we would love to learn more about your work at cosmic braini. Anto, could you share could you briefly describe your overall approach to training humanoid robots for complex tasks also factory robots? I know you have different kinds of robots. Absolutely. That's a really good question to ask actually. So as as humanity progresses from LLMs, we are getting into this space called physical AI. So I believe the next era of AI is going to be physical intelligence. and one of the biggest problems that we facing in the world is labor shortages.

So if you look at Japan or the Middle East or even the US, there are no workers in the factories and one of the solutions that we have is to use humanoid robots, right? right now the industry is divided into two. We have like legacy robotics companies like ABB, KUKA, Boston Dynamics and all that. We also have new companies like figure AI, Optimus and all that right most of these robots are very pre-programmed and do not have enough intelligence to actually do a diverse set of task and that's where we come in. We building the brain for these robots so that one robot can actually have thousand different skill sets and handle diverse amount of data sets or like environments in one go. So that's what cosmic brain is doing. Yeah. And building on that, how do you ensure that those models transfer effectively to real hardware in real planetary conditions?

Absolutely. So, so there was this breakthrough that happened in January, from Nvidia called Cosmos. So, Cosmos, the paper we believe, it's kind of like the attention is all you need paper that enabled OpenAI to build chat GPT. the what the paper did was it combined world foundation models with a physics engine and the digital twin of the robots. So at this point like in 2025 anyone out there is able to take world foundation models from large amount of video data sets of robots doing different tasks and combine with it with actual robot and make it do task in the simulation. So what we do is that we take thousand hours of real data and convert that into 100,000 hours of model data right synthetic data which will enable these robots to how to grab an object like an apple and orange in your kitchen counter to doing like complex task like in warehouses packaging in Amazon warehouses or any factories out there right so the physics is advancing as we go we also have a differential physics engine that is actually powering the simulations for actual real world robots otics as well.

Fantastic. Thank you, Anto. Jason, let's move to you. Your research, your work related to training robots and aircrafts for the Titan moon mission in Mars exploration. All that sounds so inspiring. Could you briefly describe your overall approach to training robot rovers and aircrafts for the Titan moon mission? Sure. Yeah. So, so we work on a couple different space programs. obviously of course NASA does many but my my group we've worked very closely with the Ingenuity Mars helicopter which flew 72 times on Mars in the last few years and then a lot of what I work on for the last several years is Dragonfly. So this is a fully autonomous helicopter really a relocatable lander is what we call it and it's going to be put on a rocket in 2028 launched to Saturn's moon Titan and then it will operate there for quite quite a while.

And so we use, you know, AI and machine learning a lot in kind of how we approach different problems here. And one of them kind of the specialty that I'm getting into now is optimizing the aerodynamics and maybe stability and control or the design of these vehicles using those types of methods. So one of the speakers this morning was kind of talking about different tiers of AI and there's kind of the classical machine learning surrogate modeling which is what we're doing a lot of now. So we have inputs from our simulations that we make a huge database on the NASA supercomputer and then we can train classical machine learning models to have a relationship between design variable inputs and then the output of performance. And so we're doing a lot of this to kind of accelerate how quickly we can look at thousands or maybe tens of thousands or even millions of design perturbations of the same aircraft and then come up with the best possible aircraft to achieve the mission that we need to on Titan for example.

and so right now we're doing the same thing for Mars and basically looking at how we can make a long range Mars rotorcraft where we're using the supercomputer. And so we just got 350 new H100 GPUs. so it's a small number compared to some of the the things that other folks are talking about here today. but we're having a really nice time using that to advance all the other work that we're doing at NASA. Do you use similar pipelines for both Mars and Titan? Similar pipelines? yeah they are they can be quite similar. So I mean at the end of the day right if we're talking about a helicopter or an airplane flying around the fundamental underlying physics is quite similar. there are some unique peculiarities between the two. So for example you know the Martian atmosphere is about 1% the atmospheric density of what we have here around us.

So if the air around us is 1.225 kg per meter cubed on Mars it's about 1% of that. On Titan it's about 440%. So the physics though, I mean the Navier Stokes equations that we're running in our supercomput calculations, the physics is the same and then you swap out these numbers. and so the result of this is that you will get different configurations of a helicopter or an airplane or whatever you choose to go with. It will look a little bit different on each of these locations, but at the end of the day, the process that we're using to design them, the pipeline, as you say, is is quite similar. Amazing. Thank you both. That was really enlightening. I think now we can explore a couple of common threads on GPU based acceleration and superclusters for training. Can you share with the audience any best practices from your experience?

Sure. yeah. So one of the things that I will remember from XAI so XAI had the largest 100,000 superclusters right. So when handling with a complex models like video models or simulation models which takes 10 million hours of data 20 million hours of video data right you need a lot of compute a lot of like high-end GPUs like H100s or H200s and you need a massive amount of memory bandwidth right so GPU optimization is one of the crucial things when it comes to this like so we have like data parallelization techniques or offloading the you know the load into 100,000 GPUs all at once, you know. So those are some of the techniques that we use to actually distribute the the the load of the models into a larger cluster. this is better for you know saving a ton of cost. it saves a ton of power power.

You don't require as much of power and also the models are not laggy. So when you have like lot of GPUs working together it's prompt in response. So any task that you give the models to actually perform it's on time very fast. So these are the things that Yeah. Any best practices from your side Jason? Sure. Yeah. So you know we use it I guess in a little bit of a of a unique way where actually the really computationally intensive part of what my group does is calculating the actual simulations and physics of modeling these aircraft on the other planets. So for example, if we're going to run one what we call a computational fluid dynamic simulation. So we're basically running millions and actually hundreds of millions of calculations, partial differential equations iteratively on the supercomput. And so this is really the very expensive part.

And so we've been working on this problem for quite a long time. and normally we would use CPU. So it's very typical to run this kind of problem, you know, one simulation of our aircraft in one flight condition on say for example a few thousand CPU cores. Now, more recently, we've had engineers and computer scientists that are taking these same programs that we use, but then moving them on to the GPUs. Historically, the memory size, the memory limit of the GPUs has been the big constraint there. So, it was actually previously a big challenge for us to port those over. But now with the new chips, we're finally getting a large enough memory on board of a single GPU where we can take pretty a pretty large problem and put it on a single one. And so the way we do it now is we parallelize thousands or tens of thousands of perturbations of our vehicle design across all of the different GPUs that we have available, run them all, and then when we do the actual machine learning now that we have this training data set that we've made, which you can kind of think of it as maybe some synthetic data, right, you've kind of perturbed a massive design space with all of these different potential solutions.

And then we use that to train a surrogate model. and since we have the GPUs, we can train hundreds or thousands of surrogates and then pick the best one and use this to do further optimization. so hopefully that kind of answers best practices at least for what we're doing. Thank you. Last question. Any Jason, what's one lesson you learned about bridging the gap between simulation and reality at such large scale, planetary scale? Sure. Yeah. So you know the interesting thing on on our side I would say that I was in school trained as sort of a numerical person. So I was doing these CFD simulations on the supercomputers and it's funny because then you talk to the experimental folks that at NASA as down the road in Mountain View we have one of the largest wind tunnel facilities in the entire world. Actually the largest and so the folks that run that they never believe the simulation.

The people that run the simulations they never believe the experiment. And so I'm kind of I've come to the approach where you need a little bit of everything. So even though we have all of this amazing computational resource now, we still need to validate it. And there were some talks earlier today about how you validate the output we get from these models. and I think this will always be a critical step for us to really believe what we find. Thank you. so to answer that question, so I do believe that simulations you said simulations people do not actually believe in the hardware of the stuff, right? So simulations are definitely the future because we do not have enough real world data to train the robots as of today. So we need more synthetic data and that will come from simulations, right? So one of the things that if you look into this re world right now, so you might be wondering, hey, why are there no humanoid robots launched as of today, right?

well Tesla said they will launch Optimus in 2027. figure AI is coming out with the robots in the next one year mass manufacturing but as of today you see robots advancing in hardware but not much in AI right if you ask a robot to do some particular task it can be like kind of like really basic one of the reasons being for the past 50 years the robots have been trained in a very legacy manner right so a lot of these companies for training they use hundreds of different tools so if you train the robot for in a physics engine you transfer the file to a simulation environment like Omniverse or ISXM right lot of data is getting lost in the transfer right and that's where a solution needs to take place we need to build a model that combines physics world foundation models and simulations all together with high fidelity so no data is lost that is crucial for optimizing task or skill sets for these robots and make it happen right and we believe this is the gap that is preventing the physical robots into actually coming into your home environments assisting you with laundry or you know cooking a food for you or assisting you in your factory floors.

So there is a big gap between hardware and software. So in the past two years hardware has gone super big. China has launched back flipping robots and running robots. You can see all over the place. But right now it's time for us to build the brain and advance it with AI and bridge that gap. So that's where we are at. Well, we've covered a lot in 15 minutes. Thank you so much Jason and Anto for sharing your insights on how NASA supercomputers and simulations are shaping and pushing the planetary exploration. Thank you.