# Solving the “Three Computer” Problem: Building the Cloud Infrastructure for Physical AI (Presented by Nebius)

## Abstract

Physical AI represents a fundamental shift in computing, demanding the seamless integration of three distinct workloads—high-fidelity simulation, large-scale model training, and low-latency real-world inference—which NVIDIA CEO Jensen Huang calls the "Three Computer" problem. Join this panel featuring leaders from Nebius and pioneering robotics and simulation companies as they dissect what it takes to bring this architecture to life. Moving beyond theory, the discussion will focus on the concrete infrastructure, orchestration layers (including NVIDIA OSMO and Cosmos), and pipeline strategies required to close the loop between virtual worlds and physical reality.

## AI Summary

- The panel discussed the importance of simulation, robotics, and AI in creating efficient and reliable physical AI systems, emphasizing the need for a robust infrastructure to handle data and workloads across training, simulation, and edge computing environments.
- The challenge of making simulations representative of real-world robot performance was highlighted, along with the difficulty of integrating and scaling up data from various sources, such as real-world and synthetic data, to ensure high accuracy and reliability.
- The concept of a data engine was explored, focusing on the need for physically accurate data, flexibility to adapt to new algorithms, and the ability to generate synthetic data that can be combined with real-world data to improve model training and deployment.
- The discussion also touched on the importance of continuous learning in physical AI systems, where every bit of data and information is used to improve and adapt the models, and the need for organizations to be agile and open to technological advancements.
- The panelists emphasized the critical need for security in physical AI systems, especially in manufacturing, to protect against adversarial AI and ensure the integrity of the data supply chain.
- The session concluded with a call for manufacturing leaders to start preparing for the integration of physical AI, as the technology is rapidly evolving and the early adopters will have a significant advantage.

## Transcript

**[00:00:06 – 00:00:08]** Thank you.

**[00:00:08 – 00:00:11]** Okay, guys, I encourage you.

**[00:00:11 – 00:00:13]** Join me.

**[00:00:14 – 00:00:16]** All right, hi.

**[00:00:17 – 00:00:21]** Okay, thanks everyone for joining us.

**[00:00:21 – 00:00:26]** My favorite topic, three computers, simulation, robotics.

**[00:00:26 – 00:00:29]** So we'll do some table setting to start.

**[00:00:29 – 00:00:33]** I just wanna say this is one of the best panels I could

**[00:00:33 – 00:00:36]** have imagined for this topic across like every layer of the stack.

**[00:00:36 – 00:00:39]** And so I'm excited for you guys to hear from each of these folks.

**[00:00:39 – 00:00:43]** So if you guys could please introduce yourself, your company.

**[00:00:44 – 00:00:46]** And the role you're playing in the physical AI ecosystem.

**[00:00:46 – 00:00:49]** And as you're doing it, try to articulate for us like the

**[00:00:49 – 00:00:53]** main problem you're solving within your business for your customers.

**[00:00:53 – 00:00:54]** So I'll start with you right there.

**[00:00:54 – 00:00:55]** So me, okay.

**[00:00:55 – 00:00:57]** Hey everyone, I'm Lerel.

**[00:00:57 – 00:01:00]** I'm the CEO of Assured Robot Intelligence.

**[00:01:00 – 00:01:05]** So we build frontier AI models for robots, all sorts of robots,

**[00:01:05 – 00:01:08]** most prominently humanoid form factor robots.

**[00:01:09 – 00:01:12]** So we do a lot of the AI aspects, right?

**[00:01:12 – 00:01:14]** So we don't build as much of the hardware.

**[00:01:14 – 00:01:19]** So for us, what's really important is infrastructure for training,

**[00:01:19 – 00:01:23]** infrastructure for evaluating, for understanding how good

**[00:01:23 – 00:01:26]** the models are, and then finally infrastructure for actually

**[00:01:26 – 00:01:32]** deploying these models on the robot in a highly efficient,

**[00:01:32 – 00:01:36]** highly sort of edge way, yeah.

**[00:01:37 – 00:01:40]** Hi everyone, my name is Louis Lian, it's not Steve Xie.

**[00:01:40 – 00:01:44]** Sorry for this, our founder and CEO Steve has an emergency

**[00:01:44 – 00:01:46]** that he cannot attend, so I'm replacing him.

**[00:01:46 – 00:01:51]** So I'm in charge of global partnership and strategy in

**[00:01:51 – 00:01:55]** Lightwheel, and as well I'm the classmate of Steve, so we know

**[00:01:55 – 00:01:56]** each other for many, many years.

**[00:01:56 – 00:01:58]** Both of us studied physics.

**[00:01:58 – 00:02:03]** That's why we believe that physics AI has to be successful.

**[00:02:03 – 00:02:09]** And what we do, actually we are a simulation infrastructure company.

**[00:02:09 – 00:02:10]** So we start from the solver itself.

**[00:02:10 – 00:02:13]** We build physically accurate solver.

**[00:02:13 – 00:02:18]** And then we build the sim-ready assets so that it can run,

**[00:02:18 – 00:02:21]** it can manipulate on the Isaac Seaman Newton environment.

**[00:02:21 – 00:02:23]** And then we have the evaluation system.

**[00:02:23 – 00:02:30]** We have the big industrial-grade benchmark for the robots

**[00:02:30 – 00:02:34]** so that we can test, we can orientate the research

**[00:02:34 – 00:02:36]** directions, and we can deploy.

**[00:02:36 – 00:02:40]** So all what we do is try to make simulations successful,

**[00:02:40 – 00:02:44]** and we really believe this in the coming years.

**[00:02:44 – 00:02:45]** That's right.

**[00:02:45 – 00:02:47]** My name is John Quick. I'm the CEO of Launchpad.

**[00:02:47 – 00:02:50]** What Launchpad is trying to solve is the fact that we've

**[00:02:50 – 00:02:53]** had robots for 20, 30 years and then the number of robots or the

**[00:02:53 – 00:02:57]** percent of manufacturing assembly done by robots is less than 3%.

**[00:02:57 – 00:02:58]** It's just way too expensive.

**[00:02:58 – 00:03:00]** So if you're Ford or GM, you've got robots and if

**[00:03:00 – 00:03:02]** you're everyone else, you don't.

**[00:03:02 – 00:03:05]** And in the US, 95% of factors are small businesses.

**[00:03:05 – 00:03:07]** So we need to solve that. In order to do it, you need

**[00:03:07 – 00:03:10]** to have some form of AI in order to be able to do the

**[00:03:10 – 00:03:12]** upfront design because you can't afford to make everything bespoke.

**[00:03:12 – 00:03:15]** And so we're really trying to be the Mr. Potato Head

**[00:03:15 – 00:03:18]** of robots, where one robot becomes tens of thousands of robots.

**[00:03:18 – 00:03:22]** And how do you do that effectively and affordably, which is why

**[00:03:22 – 00:03:24]** simulation and digital twins is at the heart of that.

**[00:03:24 – 00:03:26]** And you worked at an autonomous systems company that did

**[00:03:26 – 00:03:28]** simulation for that.

**[00:03:28 – 00:03:30]** Exactly.

**[00:03:30 – 00:03:32]** John and I worked together at a company called Liquid

**[00:03:32 – 00:03:36]** Robotics where we did very long-duration autonomous robots.

**[00:03:36 – 00:03:39]** Actually, the first robots that would really run autonomously

**[00:03:39 – 00:03:43]** for years at a time on their own and operate in the environments and

**[00:03:43 – 00:03:47]** so do swarming and stuff and then I was at AWS for about 10 years

**[00:03:47 – 00:03:51]** working on robotics and simulation and a bunch of other things and now

**[00:03:51 – 00:03:55]** I'm the CTO for Booz Allen where we do a lot of autonomy for drones and

**[00:03:55 – 00:03:59]** All sorts of autonomous land and sea and air vehicles and even

**[00:03:59 – 00:04:02]** space satellite desegregation and other training around those areas.

**[00:04:02 – 00:04:07]** So we do a lot of synthetics and a lot of synthetic data

**[00:04:07 – 00:04:11]** and simulations engines and a lot of 3D spatial computing which

**[00:04:11 – 00:04:13]** we used to do when I was at AWS.

**[00:04:13 – 00:04:15]** Yes, we were a little bit early though.

**[00:04:15 – 00:04:17]** We were banging our head against the wall trying to get this

**[00:04:17 – 00:04:21]** stuff to be real six years ago and now it's all happening.

**[00:04:21 – 00:04:23]** So thank you for that.

**[00:04:23 – 00:04:25]** Each of you guys are solving a pretty unique business

**[00:04:25 – 00:04:27]** problem in the industry.

**[00:04:27 – 00:04:33]** But on that path, you're all wrestling with a similar problem

**[00:04:34 – 00:04:34]** to bring your solutions to life.

**[00:04:35 – 00:04:37]** And that's this three-computer problem that we're

**[00:04:37 – 00:04:39]** all talking about.

**[00:04:39 – 00:04:43]** And so I'd like to start by trying to find the problem,

**[00:04:43 – 00:04:44]** but one click down.

**[00:04:44 – 00:04:46]** I think everyone understands.

**[00:04:47 – 00:04:49]** At a high level, this is all about, how do you orchestrate

**[00:04:49 – 00:04:53]** data and workloads across three disparate compute environments

**[00:04:53 – 00:04:56]** that are all their own systems, their own configurations?

**[00:04:56 – 00:04:58]** And those three systems are the training environment,

**[00:04:58 – 00:05:02]** the simulation environment, and inference at the edge.

**[00:05:03 – 00:05:07]** But it'd be great to hear from maybe you first, Leryl.

**[00:05:07 – 00:05:11]** Why does this problem exist from a technical perspective?

**[00:05:11 – 00:05:14]** Like, why do we have to struggle so much across these three computers?

**[00:05:14 – 00:05:17]** Maybe you can give us a bit of a sense of a day in the

**[00:05:17 – 00:05:20]** life of a developer at your company as you guys work through this.

**[00:05:20 – 00:05:23]** Yeah, so I think there are sort of

**[00:05:23 – 00:05:27]** several ways of thinking of this i think the most top

**[00:05:27 – 00:05:31]** level thing is when you think about just normal digital ai right

**[00:05:31 – 00:05:35]** everything is digital everything is on a specific layer but the moment

**[00:05:35 – 00:05:39]** you go into the physical world the fundamental harsh reality is

**[00:05:39 – 00:05:43]** that you need to get things to work on a real robot you have to get

**[00:05:43 – 00:05:46]** things to work on a real system so

**[00:05:46 – 00:05:51]** All of this three computer problem, all of this sort of challenge

**[00:05:51 – 00:05:55]** actually stems from the fact that ultimately all of this has to go

**[00:05:56 – 00:05:58]** on this physical platform, right?

**[00:05:58 – 00:06:01]** And so if you talk about like evaluations and simulation,

**[00:06:01 – 00:06:03]** why is evaluation and simulation so important?

**[00:06:03 – 00:06:07]** It's because you cannot just run every single test run on a robot.

**[00:06:07 – 00:06:08]** It's going to be too expensive.

**[00:06:08 – 00:06:10]** Why is edge compute so important?

**[00:06:10 – 00:06:13]** It's because when you're running things on a robot, you need

**[00:06:13 – 00:06:14]** it to have low latency.

**[00:06:14 – 00:06:15]** Maybe you need to

**[00:06:15 – 00:06:17]** To be quick in inference, you need to have a certain

**[00:06:17 – 00:06:20]** level of properties on it.

**[00:06:20 – 00:06:23]** So I think that's the first framing where, because of

**[00:06:23 – 00:06:31]** the fact that it is on the physical system, it requires the creation of

**[00:06:32 – 00:06:34]** specific infrastructure for that.

**[00:06:34 – 00:06:36]** Now, we can borrow a lot.

**[00:06:36 – 00:06:39]** We can borrow a lot from what is created in the digital world.

**[00:06:39 – 00:06:42]** But things are new.

**[00:06:42 – 00:06:45]** I don't know, as you said, it was maybe hard to do these

**[00:06:46 – 00:06:47]** problems six years ago.

**[00:06:47 – 00:06:50]** But now we are able to crack these problems.

**[00:06:50 – 00:06:55]** And so we are in this place where we have sort of semblance of tools.

**[00:06:55 – 00:06:58]** But those tools are still, in a way, chaotic.

**[00:06:59 – 00:07:02]** And so we have to organize these chaotic tools in a way

**[00:07:02 – 00:07:05]** to be very efficient for what we want on a system.

**[00:07:06 – 00:07:09]** So if I'm walking through a process of, let's say, one

**[00:07:09 – 00:07:11]** of our clients come to us.

**[00:07:11 – 00:07:16]** They want to get humanoid robots to do some sort of factory line work.

**[00:07:16 – 00:07:18]** I'm just giving a broad example here.

**[00:07:18 – 00:07:20]** So first we

**[00:07:20 – 00:07:22]** Take this robot, we collect data.

**[00:07:22 – 00:07:26]** We train models in this first computer, which is

**[00:07:26 – 00:07:28]** just pure training.

**[00:07:28 – 00:07:30]** Now, once you train it, you have to eval it.

**[00:07:30 – 00:07:34]** So for different runs, you have to eval and check a performance.

**[00:07:34 – 00:07:37]** So for that, we usually build either a simulation twin, like a

**[00:07:38 – 00:07:41]** digital twin of those environments.

**[00:07:41 – 00:07:44]** Or if they're simple enough, we directly evaluate on the robot.

**[00:07:44 – 00:07:47]** So that's a second piece where if we have good simulations,

**[00:07:48 – 00:07:49]** it can give us good evals.

**[00:07:50 – 00:07:53]** And the third piece is once we get certain levels of a performance,

**[00:07:53 – 00:07:55]** we then run this on the system.

**[00:07:55 – 00:07:57]** But that's not it. Once you run it on the system, you

**[00:07:57 – 00:07:59]** get feedback from that system.

**[00:07:59 – 00:08:02]** And you can take this feedback and then go all the way back

**[00:08:02 – 00:08:04]** to the first computer.

**[00:08:04 – 00:08:07]** So these three things are not separate.

**[00:08:07 – 00:08:09]** They're actually connected to each other, and they're sort

**[00:08:09 – 00:08:13]** of all looping together, where data is being shared all over the place.

**[00:08:14 – 00:08:16]** Is there one bottleneck that comes up the most in that

**[00:08:16 – 00:08:19]** loop for you guys that's the biggest pain point today?

**[00:08:19 – 00:08:23]** There are so many bottlenecks, it's hard to... Perfect,

**[00:08:23 – 00:08:24]** that's why we're here.

**[00:08:24 – 00:08:25]** What are the top three?

**[00:08:25 – 00:08:30]** Yeah, so I think first one is probably the eval sets

**[00:08:30 – 00:08:34]** being representative of the actual robot performance.

**[00:08:34 – 00:08:36]** You have to, you know, it's always hard to get the sim

**[00:08:36 – 00:08:40]** to match what the real perf is.

**[00:08:40 – 00:08:42]** The second thing is

**[00:08:42 – 00:08:45]** That feedback of data back into training, because once you

**[00:08:45 – 00:08:48]** get data from the robot, it is sort of out of distribution, usually

**[00:08:48 – 00:08:50]** from what it has been trained on.

**[00:08:50 – 00:08:53]** So just that infrastructure to actually take the data

**[00:08:53 – 00:08:58]** and train that, and making sure that the models don't fail

**[00:08:58 – 00:09:00]** from the data is pretty important.

**[00:09:00 – 00:09:01]** And then I think,

**[00:09:02 – 00:09:04]** Third piece is just like how you scale it up, actually.

**[00:09:05 – 00:09:07]** So when you're working with one customer, it's easy to

**[00:09:07 – 00:09:11]** have a few folks just monitoring how everything is going.

**[00:09:11 – 00:09:13]** But when you scale things up, when you're working with many, many

**[00:09:13 – 00:09:17]** robots across many, many tasks, how do you get the infrastructure

**[00:09:17 – 00:09:21]** where you can actually check on the jobs without having one person

**[00:09:22 – 00:09:25]** just babysitting all the robots and all the models being trained?

**[00:09:25 – 00:09:28]** Sounds like an AI problem, maybe.

**[00:09:28 – 00:09:31]** So, Steve, turning to you, one of the main outliers,

**[00:09:31 – 00:09:34]** sorry, Louise, oh my gosh, I didn't memorize my script at all.

**[00:09:34 – 00:09:36]** We're role-playing today.

**[00:09:36 – 00:09:43]** I've never met Steve, so, Louise, one of the main outliers

**[00:09:43 – 00:09:48]** here relative to training LLMs is simulation, obviously.

**[00:09:48 – 00:09:50]** But I've found that it's a pretty polarizing topic.

**[00:09:50 – 00:09:52]** Some customers don't believe in it at all.

**[00:09:52 – 00:09:54]** It's real-world data or bust.

**[00:09:54 – 00:09:58]** And so tell me about, what is at the heart of that

**[00:09:58 – 00:10:00]** polarity, do you think?

**[00:10:00 – 00:10:03]** And how do you guys address those objections to customers

**[00:10:03 – 00:10:05]** that are anti-synthetic data?

**[00:10:05 – 00:10:10]** Well, I think they are talks and they are contracts.

**[00:10:10 – 00:10:14]** So basically, seen from our contract's perspective, we

**[00:10:14 – 00:10:17]** don't see it as a debate, actually.

**[00:10:17 – 00:10:21]** We see it as a consensus that all of the industry

**[00:10:21 – 00:10:23]** is buying simulated data.

**[00:10:23 – 00:10:28]** Because today, all the top labs, like NVIDIA, of course,

**[00:10:28 – 00:10:34]** Google DeepMind, or Lab, almost every serious player in the world

**[00:10:34 – 00:10:37]** model sector are our paid clients.

**[00:10:37 – 00:10:41]** And as well for the top robotic companies like Fikr, One X,

**[00:10:41 – 00:10:44]** and many, many others, and many QN.

**[00:10:44 – 00:10:46]** So everyone is our paid clients already.

**[00:10:46 – 00:10:50]** So the question is, I think, of course, they don't use

**[00:10:50 – 00:10:51]** only simulated data.

**[00:10:52 – 00:10:54]** They use everything together, which is everything is important.

**[00:10:54 – 00:10:56]** But I think the critical bottleneck we have seen

**[00:10:56 – 00:11:00]** today is really the scalability.

**[00:11:00 – 00:11:04]** Because if I draw a comparison with autonomous driving, so

**[00:11:04 – 00:11:07]** autonomous driving, Tesla has already one million cars

**[00:11:07 – 00:11:11]** everywhere, but for any robotic company, how many they have?

**[00:11:11 – 00:11:13]** A few hundred.

**[00:11:13 – 00:11:15]** If they're lucky.

**[00:11:15 – 00:11:19]** So it's three to four out of magnitude difference, right?

**[00:11:19 – 00:11:23]** And then the robots, they are not the same, right?

**[00:11:23 – 00:11:27]** They are robots with the arms, with the pinch, with the fingers,

**[00:11:27 – 00:11:28]** and they are dogs, so they are not the same.

**[00:11:28 – 00:11:34]** So the training data, the real data from a robot is now transferable

**[00:11:34 – 00:11:36]** easily to another robot.

**[00:11:37 – 00:11:41]** That's why we believe that only the

**[00:11:41 – 00:11:45]** Robot agnostic data, they will be very important for

**[00:11:45 – 00:11:48]** the scalability, which include, of course, the simulation data, but as

**[00:11:48 – 00:11:51]** well as the human egocentric data.

**[00:11:51 – 00:11:53]** They're both very important.

**[00:11:53 – 00:11:57]** And so what we focus is really on how to scale up at this

**[00:11:57 – 00:12:00]** moment, because the data are from our clients.

**[00:12:00 – 00:12:03]** Actually, what they do, they are not only do post-train.

**[00:12:03 – 00:12:07]** A lot of them do pre-train, and they really need a lot of them.

**[00:12:07 – 00:12:12]** and um above this uh the scalability thing is we

**[00:12:12 – 00:12:16]** need certain accuracy that's why we prepare we have the silver by

**[00:12:16 – 00:12:22]** ourselves we need those data to be physically accurate so the videos

**[00:12:22 – 00:12:27]** we sell are not common videos generated by a an ai there's a

**[00:12:27 – 00:12:31]** physics behind and that's the most important thing for our clients

**[00:12:31 – 00:12:33]** we're going to dive later into this

**[00:12:34 – 00:12:37]** call it a data diet balance between real and synthetic.

**[00:12:37 – 00:12:40]** But before we go there, from a use case perspective, a

**[00:12:40 – 00:12:43]** lot of experimentation across the spectrum of use cases.

**[00:12:43 – 00:12:46]** Can you help us understand where on that spectrum has simulation

**[00:12:46 – 00:12:49]** arrived and the use cases are just mandatory at this point,

**[00:12:49 – 00:12:51]** you think, in a workflow versus

**[00:12:51 – 00:12:55]** They're a bit out there still, and we have a lot of innovation

**[00:12:55 – 00:12:57]** to do to get more scalable.

**[00:12:57 – 00:12:58]** So actually, physical AI is very hard.

**[00:12:58 – 00:13:02]** So as Laura has just said, though, there are so many use cases.

**[00:13:02 – 00:13:04]** Today, we are just at the starting point.

**[00:13:04 – 00:13:06]** For simulation, though. For simulation, yeah, exactly.

**[00:13:06 – 00:13:13]** And again, with autonomous driving, a car don't interact with anything.

**[00:13:14 – 00:13:16]** If it interacts, that crush.

**[00:13:16 – 00:13:18]** If it works, it does. We don't want any crush.

**[00:13:18 – 00:13:26]** But for robot, we need to pick up my bottle, open the capsule.

**[00:13:26 – 00:13:28]** So there is a solid thing there.

**[00:13:28 – 00:13:29]** And there is liquid.

**[00:13:29 – 00:13:31]** They're deformable.

**[00:13:31 – 00:13:35]** So actually, there are a lot of advanced things need to be tackled.

**[00:13:35 – 00:13:39]** So I would say that the complexity of a robot is probably two or

**[00:13:39 – 00:13:41]** three more magnitude more complex.

**[00:13:42 – 00:13:44]** And to address this, probably we need that amount of data.

**[00:13:44 – 00:13:49]** So we adding, multiply by the two together, we are like six

**[00:13:49 – 00:13:51]** order of magnitude lack of data.

**[00:13:51 – 00:13:57]** And we need to really go to every industry cases and that don't want

**[00:13:57 – 00:14:00]** to join is, works a lot on that.

**[00:14:01 – 00:14:02]** Because...

**[00:14:02 – 00:14:06]** As one company, we can now do every verticals, and we really

**[00:14:06 – 00:14:12]** need the companies who has already their assets, their environment,

**[00:14:12 – 00:14:18]** their way of working to tell the robots how you need the robot to

**[00:14:18 – 00:14:20]** function in your real environment.

**[00:14:20 – 00:14:25]** And those things can be transferable from a real world

**[00:14:25 – 00:14:30]** video into a simulated world and repeat it so that they can scale.

**[00:14:30 – 00:14:32]** Got it, got it. Super helpful.

**[00:14:32 – 00:14:35]** So Bill, let's turn to you.

**[00:14:35 – 00:14:39]** I've seen you talk to dozens, hundreds of customers.

**[00:14:39 – 00:14:41]** We've had so many debates about this space.

**[00:14:41 – 00:14:43]** I've always loved the way you think about this from

**[00:14:43 – 00:14:45]** a systems level, right?

**[00:14:45 – 00:14:48]** From the data layer to the infra layer to the interface layer.

**[00:14:48 – 00:14:54]** And you've always had a really clear form of thinking about

**[00:14:54 – 00:14:56]** how to bring it all together and bring it all to life, despite

**[00:14:56 – 00:14:58]** the complexity of that full stack.

**[00:14:58 – 00:15:01]** Talk to us about how Booz is thinking about that stack

**[00:15:01 – 00:15:04]** and how you guys are, what is your philosophy and strategy for

**[00:15:04 – 00:15:07]** bringing that to life and making it deployable at scale for customers?

**[00:15:07 – 00:15:09]** Yeah, so I mean, you gotta start off with the fact that

**[00:15:09 – 00:15:12]** robots are really stupid.

**[00:15:12 – 00:15:16]** People don't realize how inept they really are in the end

**[00:15:16 – 00:15:19]** and that it takes a tremendously dense model to effectively

**[00:15:19 – 00:15:21]** train a robot to do anything.

**[00:15:21 – 00:15:26]** So for example, at Amazon, if we wanted a robot to pick this up,

**[00:15:26 – 00:15:29]** and this was going back before a lot of the AI capabilities that are

**[00:15:29 – 00:15:33]** there, we would take 1.7 million images Because I've every...

**[00:15:35 – 00:15:38]** And we couldn't do that, so we would take about 30 scans

**[00:15:38 – 00:15:42]** of this at 4K, and then we'd have an artist draw in between with it,

**[00:15:42 – 00:15:46]** and then we'd have a procedure that would generate 1.7 million images.

**[00:15:46 – 00:15:48]** And then we would do a synthetic testing of 750,000 grasping

**[00:15:48 – 00:15:51]** plans on one item, and there were 3.3 million items.

**[00:15:55 – 00:15:57]** Just to put that in perspective, right?

**[00:15:57 – 00:16:02]** So now you can take this initial scan and use synthetic AI

**[00:16:02 – 00:16:08]** and generate it synthetically, those millions of individuals and

**[00:16:08 – 00:16:11]** bottles of all different shapes and sizes and those kinds of things.

**[00:16:11 – 00:16:17]** And I believe that the organization that has the most amount of

**[00:16:17 – 00:16:21]** 3D assets in their simulation will win the battle of AI, if you like.

**[00:16:21 – 00:16:24]** Have the best AI out there and the richest simulation.

**[00:16:24 – 00:16:28]** So what we focus on a lot, we work a lot in a much more complicated

**[00:16:28 – 00:16:32]** world because our robots, if you're not just picking up things,

**[00:16:32 – 00:16:36]** they're flying between trees and interacting with things and

**[00:16:36 – 00:16:39]** diving underwater and dealing on the surface of the water and also

**[00:16:39 – 00:16:42]** disaggregated satellites in space.

**[00:16:42 – 00:16:46]** So we generate a tremendous amount of synthetic data and

**[00:16:46 – 00:16:49]** we combine that with real data that we can collect.

**[00:16:49 – 00:16:52]** And as the autonomy starts to operate in the real world

**[00:16:52 – 00:16:55]** in tests, that generates more real-world data which we can use

**[00:16:55 – 00:16:57]** to generate more synthetic data.

**[00:16:57 – 00:17:01]** And we'll do, for example, the equivalent of 30 years of

**[00:17:01 – 00:17:07]** a device operating in about an hour in a big simulation on the cloud.

**[00:17:07 – 00:17:10]** Using, say, Omniverse or one of the other simulated environments.

**[00:17:10 – 00:17:14]** And then we'll use that to do a really robust AI training,

**[00:17:14 – 00:17:17]** to create an inference engine that can actually fit on the device.

**[00:17:17 – 00:17:20]** And that's another problem because a lot of the devices don't have

**[00:17:20 – 00:17:23]** enough space, wattage, and power.

**[00:17:23 – 00:17:25]** We also have the ability, which is something we'll talk

**[00:17:25 – 00:17:29]** about later with AI RAN that's coming out, to move some of

**[00:17:29 – 00:17:34]** the compute to a low latency, say, cellular device, so you can split

**[00:17:34 – 00:17:38]** the brain and have the really low latency stuff on the robot and the

**[00:17:38 – 00:17:43]** higher latency stuff on actually the communications framework.

**[00:17:43 – 00:17:46]** With GPUs there, and then even some of it back in the

**[00:17:46 – 00:17:48]** cloud if you want, right?

**[00:17:48 – 00:17:52]** So I think it's really an end-to-end solution with a

**[00:17:52 – 00:17:55]** tremendous amount of simulation and training that you have

**[00:17:55 – 00:17:58]** to do to do that, to take into advantage all the things

**[00:17:58 – 00:17:59]** that can happen in the real world.

**[00:17:59 – 00:18:03]** I mean, for us, too, we have to deal with GPS-denied environments.

**[00:18:03 – 00:18:05]** We have to deal with...

**[00:18:05 – 00:18:07]** Radio-frequency-denied environments.

**[00:18:07 – 00:18:11]** We're flying or walking or driving in fog and rain.

**[00:18:11 – 00:18:13]** There's mud.

**[00:18:13 – 00:18:16]** I mean, imagine your robots on the factory floor if the

**[00:18:16 – 00:18:18]** floor was moving like the surface of the ocean.

**[00:18:18 – 00:18:21]** That's another whole layer of complexity there, too, as well.

**[00:18:22 – 00:18:25]** So all that kind of simulation is really what's necessary

**[00:18:25 – 00:18:31]** to build a dense enough inference engine to operate the autonomy.

**[00:18:31 – 00:18:34]** So zooming out from simulation and thinking about

**[00:18:34 – 00:18:36]** The other three nodes here, you mentioned a bit about training.

**[00:18:36 – 00:18:39]** I imagine when it comes to the edge, you guys have seen

**[00:18:39 – 00:18:42]** some of the most scale complexity of really anyone, whether

**[00:18:42 – 00:18:47]** it's the drones or any kind of autonomous type of device.

**[00:18:47 – 00:18:52]** But those devices are dealing with lots of different variables,

**[00:18:52 – 00:18:55]** more so than most other robots.

**[00:18:55 – 00:18:57]** Have you guys,

**[00:18:57 – 00:18:59]** Where are you guys in cracking that code between creating

**[00:18:59 – 00:19:02]** a unified loop between all of the devices at scale and

**[00:19:02 – 00:19:05]** getting... So there's parts of it that are similar for everything,

**[00:19:05 – 00:19:09]** you know, like the basic, you know, rules of robotics and not to be

**[00:19:09 – 00:19:13]** the ones that Asimov said, I wish I could use those rules, but...

**[00:19:13 – 00:19:17]** The first rule is don't bump into things.

**[00:19:17 – 00:19:19]** The second rule is don't fall over or crash.

**[00:19:19 – 00:19:21]** And the third rule is go where I tell you.

**[00:19:21 – 00:19:24]** And if we could really do those three things reliably,

**[00:19:24 – 00:19:26]** then that's a basic component.

**[00:19:26 – 00:19:30]** And then you add more complicated things like actuators or

**[00:19:30 – 00:19:34]** mission parameters that you have to operate in.

**[00:19:35 – 00:19:38]** As you said, in a lot of cases, bumping into things

**[00:19:38 – 00:19:39]** is really a bad thing.

**[00:19:39 – 00:19:41]** You don't want to be doing that, especially if you're

**[00:19:41 – 00:19:45]** flying or you're driving at a high speed or you're underwater

**[00:19:46 – 00:19:49]** and you don't want to hit the ocean floor or anything like that.

**[00:19:49 – 00:19:51]** Satellites are flying at 17,500 miles per hour.

**[00:19:51 – 00:19:54]** Doing an intersect with another satellite that's moving at

**[00:19:54 – 00:19:58]** the same speed in the other Your direction is.

**[00:20:02 – 00:20:05]** It's not gonna work that well. It's a bit challenging, yeah, so.

**[00:20:05 – 00:20:07]** That's crazy. Yep.

**[00:20:07 – 00:20:11]** So John, turning to you, a lot of this space is very

**[00:20:11 – 00:20:13]** theoretical, experimental.

**[00:20:13 – 00:20:16]** You have live systems in the wild, and you're putting them

**[00:20:16 – 00:20:20]** in brownfield factories, where there's so many things to consider,

**[00:20:20 – 00:20:23]** and there's so many gotcha moments.

**[00:20:23 – 00:20:25]** You make contact with reality in that way.

**[00:20:25 – 00:20:28]** Talk to us a bit more about your approach and some of

**[00:20:28 – 00:20:30]** those gotcha moments you guys have encountered, how you're working

**[00:20:30 – 00:20:32]** around them, And then how you're

**[00:20:32 – 00:20:34]** Going about connecting this full loop.

**[00:20:34 – 00:20:36]** Yeah, no, absolutely.

**[00:20:36 – 00:20:38]** You know, I think the first thing I would start is, you

**[00:20:38 – 00:20:41]** know, we hired a VP engineering who used to build cars for Le Mans.

**[00:20:41 – 00:20:43]** And he always said, the race is Sunday.

**[00:20:43 – 00:20:46]** And I think a big thing when you start thinking of AI in the real

**[00:20:46 – 00:20:47]** world is that it has to deploy.

**[00:20:48 – 00:20:50]** So there's no software you can just send a package.

**[00:20:50 – 00:20:51]** Once you send it, you sent in, it's gotta work.

**[00:20:51 – 00:20:55]** So taking, you know, a robot back from the field is $5,000

**[00:20:55 – 00:20:57]** each time it travels.

**[00:20:57 – 00:21:00]** So shipping it there, shipping it back isn't an option.

**[00:21:00 – 00:21:02]** Sending a tech to sit and live there to fix it is not an option.

**[00:21:02 – 00:21:03]** You've got to have something that works. You've got to

**[00:21:03 – 00:21:05]** have something that's robust.

**[00:21:05 – 00:21:08]** And it's something that people don't understand the level

**[00:21:08 – 00:21:11]** of challenge that is, just because the penalty is so

**[00:21:11 – 00:21:12]** high for it not working.

**[00:21:12 – 00:21:14]** You know, it was interesting. We almost did a deal with

**[00:21:14 – 00:21:16]** one of the biggest three companies in the world.

**[00:21:16 – 00:21:18]** And the first thing they said is we'll take three.

**[00:21:18 – 00:21:20]** And I was like, well, what do you mean? They're like,

**[00:21:20 – 00:21:23]** one on the line, one beside it, and one in the closet.

**[00:21:23 – 00:21:25]** So if this one fails, this one goes in, we ship this back.

**[00:21:25 – 00:21:28]** And at any time, we've got multiple backups.

**[00:21:28 – 00:21:30]** But that's how they think, right? Because the downtime,

**[00:21:30 – 00:21:32]** the penalty is so high.

**[00:21:32 – 00:21:35]** And so when you think about getting it to work, it's also when you

**[00:21:35 – 00:21:37]** start thinking, well, what exactly does that mean, getting it to work?

**[00:21:37 – 00:21:40]** So if you're off by a millimeter.

**[00:21:40 – 00:21:41]** That's actually a big deal.

**[00:21:41 – 00:21:45]** So the way manufacturing works today is go here, torque,

**[00:21:45 – 00:21:47]** go here, torque, go here, torque.

**[00:21:47 – 00:21:50]** If you shift by this much, the whole thing breaks.

**[00:21:50 – 00:21:53]** And the way manufacturing is done is it's a whole bunch

**[00:21:53 – 00:21:55]** of money up front to talk to me, a whole bunch of money

**[00:21:55 – 00:21:57]** to deploy, and if anything goes wrong, a whole bunch more money.

**[00:21:57 – 00:21:59]** That's why no one does it.

**[00:21:59 – 00:22:02]** And so if you want to use AI to be able to fix all that, it

**[00:22:02 – 00:22:05]** has to be able to adjust and work, otherwise you've got a problem.

**[00:22:05 – 00:22:07]** So you're talking about vision to be able to adjust, be able

**[00:22:07 – 00:22:08]** to actually hit the screw.

**[00:22:09 – 00:22:11]** The other thing is, when you start thinking about percentages.

**[00:22:11 – 00:22:14]** So if I told you that I could screw with 95% efficiency,

**[00:22:14 – 00:22:17]** you're probably thinking, that's pretty good.

**[00:22:17 – 00:22:20]** Well, if I have a window and door that I have to have eight

**[00:22:20 – 00:22:23]** screws per window and door, and at 95%, I'm gonna have

**[00:22:23 – 00:22:28]** one wrong every three windows and doors, I'm now at a 66% throughput.

**[00:22:28 – 00:22:30]** I don't have a business.

**[00:22:30 – 00:22:31]** And I've completely screwed up that line.

**[00:22:31 – 00:22:34]** So when you look at things, 95% is not pretty good.

**[00:22:34 – 00:22:37]** It is failure. It is, I don't have a company.

**[00:22:37 – 00:22:41]** So it is 99.5% minimum just to even be at the table.

**[00:22:41 – 00:22:43]** So when you think about real world, it's getting to that

**[00:22:43 – 00:22:44]** level, it's doing it.

**[00:22:44 – 00:22:47]** And that's just controlling what you can control.

**[00:22:47 – 00:22:50]** So we have a competitor who I won't mention names on, but they

**[00:22:50 – 00:22:52]** built this really great system.

**[00:22:52 – 00:22:54]** They were all excited it went there.

**[00:22:54 – 00:22:57]** they didn't realize was it was underneath the air conditioner

**[00:22:57 – 00:23:00]** so the temperature was different and all of a sudden this beautiful

**[00:23:00 – 00:23:02]** machine didn't do anything it worked in their environment

**[00:23:02 – 00:23:04]** it worked in kind of testing and everything else but secondly

**[00:23:04 – 00:23:09]** shipped it tilt well and so that's one where you know like

**[00:23:09 – 00:23:12]** what do you even do with something like that and so it's it's really

**[00:23:12 – 00:23:15]** It's the challenge here of, I need to ship it, I need

**[00:23:15 – 00:23:18]** it to work, I need to think of all the simulations for what,

**[00:23:18 – 00:23:20]** just to make it work, I need to think of what I can't really think

**[00:23:20 – 00:23:23]** about too much just to make sure it goes, and on top of all that,

**[00:23:23 – 00:23:25]** I have to do it at the right cost.

**[00:23:25 – 00:23:27]** You know, you mentioned 1.7 million.

**[00:23:27 – 00:23:28]** I can't afford 1.7 million things.

**[00:23:28 – 00:23:30]** I'm not Amazon.

**[00:23:30 – 00:23:32]** When I think of the world and I think of simulation,

**[00:23:32 – 00:23:35]** so I have a window and door robot.

**[00:23:35 – 00:23:39]** Comes in, the AI looks at it, says, I know what color

**[00:23:39 – 00:23:40]** it is, I know what size it is.

**[00:23:40 – 00:23:43]** Moves over, automatically triggers the right lock, screws it in.

**[00:23:43 – 00:23:46]** Triggers the right latch, screws it in, comes out the other side.

**[00:23:46 – 00:23:48]** The way I'm able to operate cost effectively to be able

**[00:23:48 – 00:23:51]** to deploy that robot really affordably is because I don't

**[00:23:51 – 00:23:52]** look at everything else.

**[00:23:52 – 00:23:53]** I'm just looking for the hole.

**[00:23:53 – 00:23:56]** So once I know what it is, I know exactly what to go

**[00:23:56 – 00:23:58]** trigger, and then I know exactly where to go look for the hole.

**[00:23:58 – 00:24:01]** So I'm only calibrating this little area here within a

**[00:24:01 – 00:24:02]** one meter by one meter.

**[00:24:02 – 00:24:05]** So I don't worry about the ocean shifting or...

**[00:24:05 – 00:24:07]** Temperature is on God knows what range.

**[00:24:07 – 00:24:09]** I'm just worried about a very controlled environment, controlled

**[00:24:09 – 00:24:12]** light, controlled temperature, because all those variables

**[00:24:12 – 00:24:13]** are each failure points.

**[00:24:13 – 00:24:15]** And if I can control failure points, I can actually make

**[00:24:15 – 00:24:17]** it work, which is the only way we get things actually

**[00:24:17 – 00:24:20]** out in the field working today.

**[00:24:20 – 00:24:22]** That's awesome.

**[00:24:22 – 00:24:26]** So within that, there's all kinds of challenges to create

**[00:24:26 – 00:24:29]** just the right data for that narrow use case you mentioned.

**[00:24:29 – 00:24:32]** And you guys talk a lot about data engines.

**[00:24:32 – 00:24:35]** And going back to this data diet and this balance between

**[00:24:35 – 00:24:38]** real and synthetic, how do you guys think about data engines?

**[00:24:38 – 00:24:41]** Like, can you define a data engine for me and describe what does

**[00:24:41 – 00:24:43]** good look like for a data engine?

**[00:24:43 – 00:24:45]** Is that for me? That's for you.

**[00:24:45 – 00:24:46]** For you, Steve.

**[00:24:46 – 00:24:48]** Steve, yeah.

**[00:24:49 – 00:24:54]** I think the first thing the data engine for the physical AI is,

**[00:24:54 – 00:24:55]** you need to be physically correct.

**[00:24:55 – 00:24:58]** You need to be physically accurate.

**[00:24:58 – 00:24:59]** Why we believe this?

**[00:24:59 – 00:25:03]** Because we think that physical laws, they are

**[00:25:03 – 00:25:06]** another representation of our world, which Which means?

**[00:25:07 – 00:25:10]** Text is one of the representations, all the visuals is another

**[00:25:11 – 00:25:16]** representation, and physics is a third representation of our world.

**[00:25:16 – 00:25:17]** And what model needs?

**[00:25:18 – 00:25:22]** The model needs simply more information, more data.

**[00:25:22 – 00:25:27]** So I repeat, physics is an entirely new source of information

**[00:25:27 – 00:25:29]** that models can consume.

**[00:25:29 – 00:25:32]** And the only way to give those important and useful information

**[00:25:32 – 00:25:37]** to the model is that we need to be accurate at the first place.

**[00:25:37 – 00:25:40]** Otherwise, there's no difference between this simulated environment

**[00:25:40 – 00:25:41]** versus a normal video.

**[00:25:42 – 00:25:46]** So many of the problems or debates comes from that we are

**[00:25:46 – 00:25:53]** not yet there to be 100% accurate, but we have seen a lot of progress.

**[00:25:53 – 00:25:58]** During the past years, from the CPU-based solver to GPU-based

**[00:25:58 – 00:26:02]** solver, and we are working a lot, partnered with NVIDIA

**[00:26:02 – 00:26:06]** on the new solver of Newton to make it even more accurate

**[00:26:06 – 00:26:09]** and even more scalable.

**[00:26:09 – 00:26:12]** So actually, we are trying to bring a lot of information

**[00:26:12 – 00:26:15]** to the new models.

**[00:26:15 – 00:26:16]** with the physical law.

**[00:26:16 – 00:26:19]** And this is what we are trying to do.

**[00:26:19 – 00:26:22]** And we think that this is the right way.

**[00:26:22 – 00:26:28]** And this is the first factor about the quality and the scalability.

**[00:26:28 – 00:26:30]** And second thing, I think we need a pipeline.

**[00:26:30 – 00:26:32]** And as you have said, actually, there are a lot of things

**[00:26:32 – 00:26:34]** flying around there.

**[00:26:34 – 00:26:36]** These tools are not good at, yeah.

**[00:26:36 – 00:26:40]** So actually, we need a mature industry-grade pipeline so that

**[00:26:40 – 00:26:44]** we can feed enough of this data to the models, which is, today,

**[00:26:44 – 00:26:48]** I think we are one of the pioneers in this industry trying to do this.

**[00:26:48 – 00:26:52]** So actually, where the data comes from, probably you don't know.

**[00:26:52 – 00:26:53]** We are a simulation-focused company.

**[00:26:53 – 00:26:58]** However, we are the biggest human, egocentric data supplier

**[00:26:58 – 00:27:02]** of this world, which means we have a lot of people helping us to

**[00:27:02 – 00:27:05]** film data everywhere in the world.

**[00:27:05 – 00:27:06]** And why we do this?

**[00:27:06 – 00:27:10]** Because with all this data, of course we sell them, but

**[00:27:10 – 00:27:14]** more importantly is that those data will become our evaluation

**[00:27:14 – 00:27:19]** benchmarks, will become our simulation environments, and some

**[00:27:19 – 00:27:23]** of them will even become the assets with our capabilities inside.

**[00:27:23 – 00:27:29]** And so actually we are establishing a scalable world, simulation-ready

**[00:27:29 – 00:27:31]** world with all this data.

**[00:27:31 – 00:27:33]** And this is one source of information.

**[00:27:33 – 00:27:36]** The other source, resource, it comes from the CAD, which is,

**[00:27:36 – 00:27:39]** I think John knows a lot about it.

**[00:27:39 – 00:27:43]** So many of everything we, today, we have a CAD, we

**[00:27:43 – 00:27:44]** designed it with a plan.

**[00:27:44 – 00:27:49]** So small from the bottle, large like car or a plane,

**[00:27:50 – 00:27:54]** they all already have the assets, the CAD, but this data are not.

**[00:27:54 – 00:27:57]** consumable for the robots.

**[00:27:57 – 00:28:00]** So internally, we have a very strong pipeline to convert those

**[00:28:00 – 00:28:04]** CAD files into OpenUSD, and we map those physics insights so that they

**[00:28:04 – 00:28:09]** can be SIM-ready assets, and we have put trajectories, environment,

**[00:28:09 – 00:28:11]** everything, so that we prepare.

**[00:28:11 – 00:28:16]** The sim ready data pipeline so that you can train your model directly

**[00:28:16 – 00:28:20]** So I think both of things are if we can do the both things, right?

**[00:28:21 – 00:28:22]** We have a good data engine.

**[00:28:22 – 00:28:29]** I get so let's say synthetically biased For you as a humanoid

**[00:28:29 – 00:28:31]** and model builder

**[00:28:31 – 00:28:33]** You're thinking about all kinds of data.

**[00:28:33 – 00:28:36]** How would your definition of a good data engine compare to that?

**[00:28:36 – 00:28:40]** And then also talk to us about how that engine's fitting

**[00:28:40 – 00:28:42]** into the broader architecture of the learning loop that

**[00:28:42 – 00:28:43]** you guys are building.

**[00:28:43 – 00:28:50]** VIKAS YADAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVANDAVAN

**[00:28:50 – 00:28:53]** I think there's a question about just the science of

**[00:28:54 – 00:28:56]** physical AI itself.

**[00:28:56 – 00:28:57]** And what I mean by this, right?

**[00:28:57 – 00:29:02]** So if you ask anyone today, okay, how do you do language modeling?

**[00:29:02 – 00:29:05]** They'll be like, just take this large transformer model and feed

**[00:29:05 – 00:29:07]** it data and magic it works, right?

**[00:29:07 – 00:29:09]** Now everyone says that this is the right answer.

**[00:29:09 – 00:29:13]** But if you go back like 10 years ago, maybe 15, 20 years ago.

**[00:29:13 – 00:29:16]** everyone in this language space was doing different things

**[00:29:16 – 00:29:19]** some people were doing symbolic some people were doing hierarchical

**[00:29:19 – 00:29:22]** models even in these very learning based methods there are

**[00:29:22 – 00:29:24]** so many different options out there

**[00:29:24 – 00:29:29]** I think we are closer to that environment rather than this

**[00:29:29 – 00:29:33]** environment where there's one solution that everyone agrees on,

**[00:29:33 – 00:29:37]** which is why one of your earlier questions of simulation versus

**[00:29:37 – 00:29:42]** real data versus what type of architecture should you use, what

**[00:29:42 – 00:29:45]** sort of model should you use, are you going to train them end-to-end,

**[00:29:45 – 00:29:47]** are you going to fine-tune them?

**[00:29:47 – 00:29:49]** There are all these questions which are still unanswered.

**[00:29:49 – 00:29:52]** So I want to level set in that sense where

**[00:29:53 – 00:29:57]** There is no one answer that if you ask roboticists, they'll

**[00:29:57 – 00:29:58]** all say one answer.

**[00:29:58 – 00:30:00]** I don't think that's there right now.

**[00:30:00 – 00:30:03]** So everyone's looking at different pieces of it.

**[00:30:03 – 00:30:06]** And now from what we are seeing, some algorithms

**[00:30:06 – 00:30:10]** work well for certain cases, other algorithms work well

**[00:30:10 – 00:30:12]** for some other cases, right?

**[00:30:12 – 00:30:15]** And just as an example, for locomotion, if you see any

**[00:30:15 – 00:30:18]** of the humanoids walking around, jumping around, doing back

**[00:30:18 – 00:30:22]** flips, a lot of them are trained almost entirely in simulation with

**[00:30:22 – 00:30:27]** some amount of real data to keep it close to looking human-like.

**[00:30:27 – 00:30:31]** If you look at a lot of the manipulation, the arms and hands

**[00:30:31 – 00:30:37]** doing more dexterous behaviors, that's mostly trained on real data.

**[00:30:37 – 00:30:42]** So there's a wide variety of types of data and wide variety

**[00:30:42 – 00:30:44]** of use cases for this type of data.

**[00:30:44 – 00:30:48]** So with that, when you talk about data engine, it actually

**[00:30:48 – 00:30:50]** depends on what type of task you're trying to solve.

**[00:30:51 – 00:30:56]** So when I'm working with like, let's say a data vendor or

**[00:30:56 – 00:30:59]** even with our own data ops team, we need to make sure

**[00:30:59 – 00:31:03]** that for the tasks that we are trying to collect data, for each

**[00:31:03 – 00:31:06]** of them, you are hitting certain,

**[00:31:06 – 00:31:09]** sort of success rates right so example if you want to get

**[00:31:09 – 00:31:12]** motion capture data for locomotion you want to make sure that they're

**[00:31:12 – 00:31:15]** hitting certain accuracy rates on that if you're trying to get real

**[00:31:15 – 00:31:19]** world data you want to get certain success rates of how often did the

**[00:31:19 – 00:31:23]** person actually successfully do the task right and so there's there's

**[00:31:23 – 00:31:28]** that piece of things um and so

**[00:31:28 – 00:31:31]** Apart from just the accuracy piece, the other question is flexibility.

**[00:31:32 – 00:31:36]** You need this data engine to be flexible to new algorithms

**[00:31:36 – 00:31:40]** because, as I mentioned, the science is evolving.

**[00:31:40 – 00:31:42]** Every year, the models are changing.

**[00:31:43 – 00:31:45]** It's changing very rapidly.

**[00:31:45 – 00:31:48]** So however you get this data, however it's annotated, however

**[00:31:48 – 00:31:53]** it's stored and processed, it needs to be done in a way that you can

**[00:31:53 – 00:31:57]** easily run it with new algorithms.

**[00:31:57 – 00:31:59]** So I think there's that aspect of usability as well, which

**[00:31:59 – 00:32:02]** I think I want to make sure that when I get data, I can

**[00:32:02 – 00:32:04]** actually, it's actually flexible.

**[00:32:04 – 00:32:07]** Yeah, yeah. That's a great nuanced approach, a way of

**[00:32:07 – 00:32:07]** thinking about it.

**[00:32:07 – 00:32:10]** I'm gonna go off script here a little bit.

**[00:32:11 – 00:32:14]** We talked a lot about simulation and inference at the edge.

**[00:32:14 – 00:32:17]** You and I have talked about the idea of using inference

**[00:32:17 – 00:32:20]** for post-training and RL in some variety of ways.

**[00:32:20 – 00:32:22]** Can you speak a little bit about y'all's thesis?

**[00:32:22 – 00:32:23]** Yeah, yeah.

**[00:32:23 – 00:32:28]** So if you think about the general framework of training

**[00:32:28 – 00:32:31]** any AI system, it's like you have this training phase.

**[00:32:31 – 00:32:34]** You get a bunch of data, train the system, and then you have

**[00:32:34 – 00:32:35]** a deployment phase, right?

**[00:32:36 – 00:32:39]** And usually the system is frozen in the deployment phase.

**[00:32:39 – 00:32:41]** But if you think about more

**[00:32:41 – 00:32:44]** biological way of doing learning, you're always learning.

**[00:32:44 – 00:32:46]** You're never like, oh, you're frozen.

**[00:32:46 – 00:32:48]** You cannot learn anything anymore.

**[00:32:48 – 00:32:50]** You just have to be deployed, right?

**[00:32:50 – 00:32:54]** And so if you really embrace this idea that the system

**[00:32:54 – 00:32:58]** can always learn, the system can always assimilate new information,

**[00:32:58 – 00:33:02]** That is a fundamentally new way of doing learning.

**[00:33:02 – 00:33:05]** A fundamentally new infrastructure is needed to allow that

**[00:33:05 – 00:33:07]** type of learning.

**[00:33:07 – 00:33:10]** And so if you think about physical AI systems, I really think that

**[00:33:10 – 00:33:14]** we have to go towards physical AI systems that are always learning.

**[00:33:14 – 00:33:18]** Every single bit of data, every single pixel of information

**[00:33:18 – 00:33:21]** it's seeing, it's always learning from that data, right?

**[00:33:21 – 00:33:22]** Now, how do we do it?

**[00:33:22 – 00:33:23]** How do we do it scalably?

**[00:33:23 – 00:33:25]** How do we do it with inference compute?

**[00:33:25 – 00:33:28]** How do we interface this inference compute with larger, more

**[00:33:28 – 00:33:31]** heavier offline compute?

**[00:33:31 – 00:33:34]** These are all open questions that we are actively working on.

**[00:33:34 – 00:33:36]** That's awesome.

**[00:33:36 – 00:33:39]** So we're going to do a quick lightning round here to wrap.

**[00:33:39 – 00:33:41]** But we're going to leave a few minutes at the end for questions.

**[00:33:41 – 00:33:43]** So if you all want to start thinking of some questions,

**[00:33:43 – 00:33:46]** if you have any, now's a good time.

**[00:33:46 – 00:33:49]** So I like to end things.

**[00:33:49 – 00:33:56]** Both in person and on my podcast with not answers but a question.

**[00:33:56 – 00:33:58]** And as we think about this next five, 10 years, things

**[00:33:58 – 00:33:59]** are gonna get really weird.

**[00:34:00 – 00:34:02]** We're at the exponential takeoff point.

**[00:34:02 – 00:34:03]** Going to get weird?

**[00:34:03 – 00:34:05]** It's getting weird now.

**[00:34:05 – 00:34:10]** It's getting weird. So I'm curious what you got, and

**[00:34:10 – 00:34:14]** for each of you, what do you think is the most important question that

**[00:34:14 – 00:34:18]** leaders should be asking themselves and their teams as we enter this?

**[00:34:18 – 00:34:21]** Brave new world. And we can go down the line, maybe

**[00:34:21 – 00:34:24]** start with you, Bill, down the line, and do a quick.

**[00:34:24 – 00:34:28]** I think, you know, as we, for any AI right now, there's

**[00:34:28 – 00:34:30]** been a tremendous amount of data that falls on the floor

**[00:34:31 – 00:34:32]** that you don't keep.

**[00:34:32 – 00:34:35]** And so one of the most important things you should be doing

**[00:34:35 – 00:34:38]** is what's the most important Data you need to be keeping.

**[00:34:38 – 00:34:41]** Like, what are you going to need to train these inference

**[00:34:41 – 00:34:42]** models and things like that?

**[00:34:42 – 00:34:46]** I mean, from good physical AI data or good physical data, you

**[00:34:46 – 00:34:48]** can generate synthetic data, right?

**[00:34:48 – 00:34:52]** But if you don't have enough good real data to create enough

**[00:34:52 – 00:34:57]** richness to mix that with your synthetic data and data generation,

**[00:34:57 – 00:35:01]** you're not going to have a good piece of autonomy, right?

**[00:35:02 – 00:35:05]** And I think I see so many organizations that could collect

**[00:35:05 – 00:35:08]** data not doing it, right?

**[00:35:08 – 00:35:09]** So don't throw data away.

**[00:35:09 – 00:35:11]** That would be one piece, I think.

**[00:35:11 – 00:35:15]** I think the other piece is also really remember about security.

**[00:35:16 – 00:35:20]** I think there's a lot of challenges with some of the adversarial

**[00:35:20 – 00:35:22]** AI you're going to see.

**[00:35:22 – 00:35:25]** And things like that that can corrupt the supply chain for

**[00:35:25 – 00:35:28]** a model and those kinds of things.

**[00:35:28 – 00:35:30]** And then last but not least, you've got to think about

**[00:35:30 – 00:35:32]** space wattage and power, right?

**[00:35:32 – 00:35:35]** Because on a robot, you're going to have to have enough

**[00:35:35 – 00:35:37]** space wattage and power to run your inference engine.

**[00:35:38 – 00:35:41]** And so really understanding the minimal sets you can get

**[00:35:41 – 00:35:45]** there for your deployment platform is going to be important.

**[00:35:45 – 00:35:46]** Those are great.

**[00:35:46 – 00:35:48]** What about you, John? I've got two.

**[00:35:48 – 00:35:51]** You went with the Asimov quote, so I'll do war games, which is

**[00:35:51 – 00:35:53]** the only way to win is to not play.

**[00:35:53 – 00:35:55]** And what I mean by that is one of the most important things

**[00:35:56 – 00:35:57]** in simulation is knowing what you should be doing and what you should

**[00:35:57 – 00:36:00]** not be doing, because there's infinite things you could be doing

**[00:36:00 – 00:36:01]** which involves infinite computing.

**[00:36:01 – 00:36:03]** So I'll just give the one specific example of cables.

**[00:36:03 – 00:36:05]** Cables can fall in an infinite number of directions.

**[00:36:05 – 00:36:07]** The physics just hurts my brain even thinking of it.

**[00:36:07 – 00:36:10]** So if you know enough to say forget cables, let's just put that

**[00:36:10 – 00:36:11]** aside and focus on what we can do.

**[00:36:11 – 00:36:14]** The amount of time, money, and energy, whether it's in

**[00:36:14 – 00:36:16]** your simulation, digital twin, goes down exponentially.

**[00:36:16 – 00:36:19]** So how can we frame problems in a way that's solvable now

**[00:36:19 – 00:36:21]** and just say, okay, here's the box we're gonna play in

**[00:36:21 – 00:36:23]** versus what we're not gonna play.

**[00:36:23 – 00:36:26]** And then the second half of that being open to change.

**[00:36:26 – 00:36:29]** The tools, the compute, the technology, the physics is

**[00:36:29 – 00:36:31]** evolving so rapidly.

**[00:36:31 – 00:36:33]** None of this was possible two years ago.

**[00:36:33 – 00:36:37]** What was possible today versus 12 months ago is incredible.

**[00:36:37 – 00:36:38]** What's going to be possible 12 months from now?

**[00:36:38 – 00:36:41]** So what is it we do now versus what do we not do?

**[00:36:41 – 00:36:43]** And then also keeping a finger on how that evolves so I know

**[00:36:43 – 00:36:44]** what to add to that.

**[00:36:44 – 00:36:46]** That'd be Model Agile.

**[00:36:46 – 00:36:47]** Yeah.

**[00:36:47 – 00:36:50]** This morning, Jason's speech has resonated a lot.

**[00:36:50 – 00:36:54]** He said, every company should have a cloud strategy.

**[00:36:55 – 00:36:56]** So yeah.

**[00:36:56 – 00:36:59]** So actually, this is something we have been thinking.

**[00:36:59 – 00:37:05]** Every day, both at the top level, how our company needs to evolve,

**[00:37:05 – 00:37:09]** and at the individual level, how we can have our employees to success.

**[00:37:09 – 00:37:12]** So actually we are in that company level, we have already

**[00:37:12 – 00:37:16]** a light wheel cloud that automatically generates

**[00:37:16 – 00:37:18]** simulated data and benchmarks.

**[00:37:18 – 00:37:21]** It's not perfect, but we are trying to, we are there.

**[00:37:21 – 00:37:22]** It is not safe enough.

**[00:37:22 – 00:37:24]** We need other things to do it.

**[00:37:24 – 00:37:27]** And in personal level, in employee level, so actually

**[00:37:27 – 00:37:32]** we are thinking how to empower all of our employees to be a superman.

**[00:37:32 – 00:37:36]** So each one of us need to be a superman with the cloud.

**[00:37:36 – 00:37:39]** And I think this is probably the biggest question we are

**[00:37:39 – 00:37:43]** thinking every day as the manager, is that how we organize

**[00:37:43 – 00:37:47]** a workforce, an organization which is cloud-native, not

**[00:37:48 – 00:37:49]** only AI-native, but cloud-native.

**[00:37:49 – 00:37:55]** How to have agents and humans working together in an efficient

**[00:37:55 – 00:37:58]** and respectful way.

**[00:37:59 – 00:38:02]** So I think a lot of the good answers have been taken.

**[00:38:03 – 00:38:04]** Sorry for this.

**[00:38:04 – 00:38:08]** How about a more like philosophical question about just building a

**[00:38:08 – 00:38:11]** business and being a human in this?

**[00:38:11 – 00:38:15]** So one of the things that we have seen working with large

**[00:38:15 – 00:38:20]** manufacturers, actually, you know, we are all in this AI world, right?

**[00:38:20 – 00:38:22]** Everything is moving so fast, so quickly.

**[00:38:22 – 00:38:24]** There's this exponential curve.

**[00:38:24 – 00:38:26]** And then you go to like,

**[00:38:26 – 00:38:28]** In Michigan, you go to the auto industry there.

**[00:38:28 – 00:38:32]** You go to just plants where they are making things that

**[00:38:32 – 00:38:34]** we use every day.

**[00:38:34 – 00:38:37]** The pace and velocity in which you can make change in the

**[00:38:38 – 00:38:41]** physical world is much slower than the pace and velocity

**[00:38:41 – 00:38:44]** that you can make change in AI.

**[00:38:44 – 00:38:49]** So I think maybe you asked a question about leadership.

**[00:38:49 – 00:38:53]** So I think leadership on the manufacturing side,

**[00:38:53 – 00:38:56]** The key question is, you know, things are not ready yet, right?

**[00:38:56 – 00:38:59]** You think about number of tokens being used for physical

**[00:38:59 – 00:39:03]** AI is basically zero compared to digital AI, right?

**[00:39:03 – 00:39:05]** I think everyone can sort of agree that, right?

**[00:39:05 – 00:39:06]** But we know that this is going to increase.

**[00:39:06 – 00:39:11]** And so if you're like a leadership on the manufacturing side,

**[00:39:11 – 00:39:15]** It's very hard, it takes time to make progress, it takes time to

**[00:39:15 – 00:39:18]** integrate new sorts of technology.

**[00:39:18 – 00:39:21]** So how do you prepare for this new world?

**[00:39:21 – 00:39:25]** Because if you start preparing now, you will have this advantage

**[00:39:25 – 00:39:27]** in being one of the first to actually deploy this in.

**[00:39:27 – 00:39:30]** So I think that's the, if I was in manufacturing, I

**[00:39:30 – 00:39:34]** think this is the best opportunity right now, because very few leaders

**[00:39:34 – 00:39:36]** are actually thinking about it.

**[00:39:36 – 00:39:37]** That's awesome.

**[00:39:37 – 00:39:40]** All right, we're getting a time's up sign, but I see one minute here,

**[00:39:40 – 00:39:41]** so I'm gonna go for the one minute.

**[00:39:41 – 00:39:43]** Who has one question? I see your hand first.

**[00:39:43 – 00:39:47]** So I know we talk a lot about physical AI needing a lot

**[00:39:47 – 00:39:47]** more data, right?

**[00:39:47 – 00:39:49]** And we compare physical AI to the digital chat box and

**[00:39:49 – 00:39:50]** we compare it to the real world.

**[00:39:51 – 00:39:59]** You also were saying there's a multiple outcome.

**[00:39:59 – 00:40:04]** So we talk about it's a multiple of like complication going from

**[00:40:04 – 00:40:06]** chatbots to automobiles to robots.

**[00:40:06 – 00:40:08]** So we need more and more data just for robots.

**[00:40:09 – 00:40:11]** So if we're spending billions of dollars just training the

**[00:40:11 – 00:40:15]** data for LLMs, how much money do we need for physical AI

**[00:40:15 – 00:40:21]** and is that actually going to happen in a short time period?

**[00:40:23 – 00:40:26]** Look, I think the important thing is why reinvent the wheel.

**[00:40:26 – 00:40:28]** And so I think there's different versions and different people

**[00:40:28 – 00:40:30]** taking different approaches, as you mentioned.

**[00:40:30 – 00:40:32]** So some people are saying, we need to move to a vision-centric

**[00:40:32 – 00:40:34]** world and let's redo everything and start with vision.

**[00:40:34 – 00:40:36]** And there's others who are saying, let's take the LLMs and

**[00:40:36 – 00:40:39]** how do we build on that to leverage the existing infrastructure?

**[00:40:39 – 00:40:40]** I know which one I want to do because I'm not sitting

**[00:40:40 – 00:40:43]** here on multi-billions of dollars.

**[00:40:43 – 00:40:46]** yet hopefully but but that's where i think you need to be

**[00:40:46 – 00:40:48]** looking at it and i think there's a world where we can do physical

**[00:40:48 – 00:40:51]** a building off the existing bits and doing a very intelligent

**[00:40:51 – 00:40:55]** and thoughtful way yeah maybe just one sentence to add simulation make

**[00:40:55 – 00:41:02]** it much cheaper in a lot cool all right guys thank you so much we're

**[00:41:02 – 00:41:05]** happy to chat afterwards and thank you to the amazing on the panel.