I'm excited to share this exclusive

robotics interview with the investing

community. Most people still think

Nvidia only powers AI in data centers,

but you're about to get an inside look

at how they're putting that same stack

into robots and physical AI. I'm joined

by Spencer Hang, product lead for

robotic software at NVIDIA. Spencer

spent the last four years scaling some

of the most advanced robotic systems on

the planet. And he had some surprising

things to say about where physical AI is

headed next and how soon it could

happen. But that's just one of the many

technologies I'll be covering live at

GTC next week. GTC is Nvidia's massive

AI conference, showcasing the biggest

breakthroughs in everything from

robotics and self-driving cars to AI

agents and the chips that power them.

They have tons of sessions on robotics

with speakers from Nvidia, Agility, and

even Tesla. And anyone who signs up for

a free online session at GTC with my

link can enter to win an Nvidia RTX 5090

graphics card. Just attend any session,

take a screenshot as proof, and send it

to me after the conference using the

links below. GTC should be on every

investor's radar and so should Nvidia's

ecosystem for physical AI because the

next chat GPT moment won't be on your

screen. It'll be robots bringing AI into

the real world. Your time is valuable,

so let's get right into it. When I think

about the whole robotics industry, I

naively think about just building

humanoid robots, but there's obviously a

lot more to that. And you know, Jensen

announced a lot of really interesting

things about robotics during the keynote

earlier this week. So, can you walk us

through Nvidia's approach to robotics at

a high level?

>> Sure. So, you've you've heard Jensen

talk about the three computer solution.

And for robotics, the way that we think

of it is you need uh you need a computer

that allows you to train the brain,

>> right? So, this is something like a DGX.

You're training your video your your

vision language action model. You're

training up your your base model like a

VLM. Um, basically any model uh that

you're going to be using for cognition

is likely going to be trained on that

DJX. Yeah. But you still need to be able

to put that into a stack and test it

inside of a simulation or more

importantly for for humanoids and these

more autonomous skills. You want to be

able to train a skill in simulation. So

we need a computer to simulate the

world. And so you have a computer that's

that's there to train the brain, a

computer that simulates the real world.

And then we have a third computer which

is actually deployed in the real world.

And that would be IGX and AGX of Jetson.

And that gives you everything from the

brain to the body uh to the physical

apparatus inside of the world. And so

that's our three computer stack.

>> So the first one is really about

training the AI model, right? The second

one is really about making sure that

model gets as much practice as it can in

a digital world before it goes into a

physical world.

>> Yeah. And that simulated world could be

used not only for training but for

evaluation. So when you train a model uh

you know an LLM or these typical uh ML

models um you train it and then you

evaluate it. And with a simul with a

policy like a robotic policy, it's not

quite the same because you have to

interact with an environment and the

environment has to react back, right? I

poke something, it has to do something.

Uh, and so you need a simulated

environment. It's not necessarily just

uh am I classifying it as a cat

correctly. And so because of that, you

need to test it inside of something

that's a proxy to the real world. Now,

because we already have that proxy for

the real world called Omniverse, we can

also generate tons of synthetic data.

And so synthetic data compensates for

the lack of real data. And what we mean

by this for physical AI when we say that

we don't have lots of real data um LLM

started with the compendium of knowledge

that humans have written over the last

couple centuries we've spent most of our

life trying to make sure that we can

instill our knowledge for the next

generation. And so it was already there

for us to to start combing through and

turning into these these language models

that you know eventually turned into

chatbt. For physical AI we don't have

the same information for contact data.

We don't know how to we haven't captured

what is it like when you take a rigid

body like a finger um like a bone or or

or a metal hand and interact with

something very very soft. That

interaction that data doesn't exist

>> and that implies that sorry not to

interrupt you but that implies that um

the video data out there isn't enough

like

>> it's not yeah so what video data gives

us and the reason why reasoning is so

important this year is it's actually if

you think of it the video models were

trained to understand semantics. How

does the world work? How do things inter

how do they relate to each other? When I

think of if I ask you to build a

kitchen, you're not going to put, you

know, a 4x4 inside of the kitchen.

You're not going to put uh a a chair on

top of the table. You're not putting a

cutting board on the floor. You know,

semantically, there's places where these

objects are supposed to live in relation

to the environment that they're in. So,

what what the video model gives robotics

is the ability to have this cognitive

reasoning. It gives you semantic

reasoning. It gives you the ability to

understand and interpret the world. But

what it doesn't do is tell you how the

world is going to interact when you

start interacting with it.

>> That physical data is where the gap is.

That's that gap. And that's why we call

it physical AI is when you're trying to

start interacting with the world, those

reactions also matter. And so we need to

have lots and lots of data of how to

interact with the world. Otherwise, you

might grab an egg at the same strength

you would grab a baseball because

otherwise you have no you have no

interpretation. They're both balls,

right? They're both spherical objects.

But the materials themselves change. And

so we need that. That's why we need to

use SE simulation like Omniverse or or

even Cosmos as a world model.

>> How do you determine like when data is

when simulated data is good enough?

>> Sure. That's um that's kind of the

million-dollar question. So simulated

data synthetic data in general is is

more of an art than a science. And the

reason I say that is because going back

to LLM, when we have a corpus of data,

we could do things like data data

engineering, feature extraction. These

were because we had real data. We

understood okay once we have a data set

how can we start uh analyzing this data

set and saying well there's certain

characteristics and you know we call

them features that don't necessarily

have any impact perceptive impact on the

model. Um if that's the case then why

include it in the in the training data.

And so we were able to actually start

engineering the data and and creating a

good corpus of of training data that

results in a really well-trained model.

For physical AI we're lacking that. You

have lots of physical data that we can

collect but even real data we're not

sure what is good data or bad data. So

for instance, all of last year, you saw

lots of people doing tele operation to

uh open drawers and and you know handle

different objects inside of like kitchen

environments or various industrial

environments. And what they were trying

to figure out is one, how do I capture a

human demonstration that's clean enough

so that way I can train a policy? When I

say clean, I mean uh humans are

imperfect. And when we grab something,

if it's a demonstration, you want to do

it perfectly because you want the robot,

you don't want them to train off of bad

demonstrations. You want good

demonstrations for the most part. And so

what is a good demonstration also kind

of compounds once you get to synthetic

data because a good demonstration may

visually look good but it might actually

not improve the model itself because it

might be looking for different features

that aren't necessarily included in the

dimensions of that data. And so there's

a lot of um open questions on what types

of data dimensions do we need inside of

this data and what types of modalities

do we need? Video, visual, contact,

action. And so we capture all sorts of

data that's not just text or video

visual anymore. It's action data, the

motions itself, it's contact data. When

I grab something, what what exactly? And

so these are all things that that we're

learning right now.

>> That's exciting.

>> Um I'd love to work through a specific

example just to understand the endto-end

chain.

>> Sure.

>> So there's an awesome robot right down

the hall from us that's doing spinal

surgery. So, can we start with the AI

model training, walk through that piece,

then walk through how it would work in

simulation,

>> and then walk through what that would

mean for the physical robot in the final

account.

>> Sure. Um, spinal surgery is a a

challenging one. It's a good one to

choose. The the reason is because it's

rigid soft body.

>> Okay.

>> And so, what I'm going to describe is

not fully available today. It's it's

just going to describe the path that

we're on and the journey that we're

going to take.

>> Are those paths different depending on

the problem set? Sorry. I'm like, so

surgery versus industrial versus

>> warehousing, you know what I mean? This

is all very different.

>> It is. It is different, but don't

imagine it as the verticals. Imagine it

as the physical the physical problems.

Okay? So, for instance, I could be in an

industrial warehouse and I could be

picking up just boxes. And so, as long

as they're relatively they're rigid

boxes and they're nor you know, they're

they're known shape, known materials,

these are things that we could handle

today. If you want to do cable

management or wiring inside of a car,

that becomes much more difficult. But

that cable management and wiring is very

similar to threading needles and and

sutures for healthcare. Right? So if you

think of of it in the physical

characteristics of the problem, then the

verticals don't matter as much. And so

what we're trying to do is is figure out

between each vertical where are the

overlaps between the near-term problems.

So that way we can meet as many of the

customer needs as possible and because

we have to solve for it eventually. It's

just where do you start taking chunks?

And so, you know, to go end to end, we

could say, uh, let's let's start with

surgery. And I'll give you I'll give you

surgery, but I'm I'm going to pair it

right next to pick and place. Okay,

they're very very similar in a sense.

Um, because you're using some apparatus

to do some type of manipulation. So, the

first thing that you would do is I need

to understand what the task is. If it's

suturing or if it's picking up a box,

um, you're either going to be training a

policy and simulation where you're

fairly certain that you can train this

just, you know, off of behavior cloning

where I capture demonstrations. you

know, I watch I I put sensors on a

human. They teleop the robot a couple of

times and then we can generate more of

those actions and then train a policy

from that. So, we might do that for

either one of these. So, if you're

working with um with inner body, for

instance, it's all squishy things.

>> You know, they're it's not crunchy, it's

all squishy, and and it's not hard and

metallic. And I mean, maybe there's a

lot of metal in there depending on who

you are. Um, but the squishiness factor

makes it much more difficult because

when you're interacting with a, you

know, you take a probe and you put it

into a squishy thing and you kind of

move it around, it's elastic. And so how

it works and how it manip how it, you

know, interacts with the ob the the tool

itself really matters to the surgeon.

And so the simulation has to be

extremely high fidelity for that.

>> In the case of picking place, if I were

to choose just a box,

>> it's actually relatively easy cuz I'm

just taking two rigid bodies. I'm

grabbing a box and so it's relatively

easier in that sense.

>> Sure. But the process for training it is

basically the same. And so as long as

the technology catches up, so once we

have um proper physics simulation for

these really elastic soft bodies or

cable or you know different types of

rope, like everything's kind of a thread

in a lot of ways. Um if you were to

catch up, then the process staying the

same, the technology basically gives you

unblocks, you know, new uh new skills,

new areas and domains that you can start

applying. And this is why the fidelity

of the simulation matters so much I

assume because you know the the more

fine the action the more accurately you

need to capture just so much different

things about the data right from the

>> yeah sim the the goal is that if

simulation can be as close to reality as

possible then we could likely turn you

leveraging agents we could automate this

data generation process. So imagine for

a minute that you do a demonstration in

the real world and I can put it into a

pipeline that takes that one

demonstration and then turns it into

thousands of different data, you know,

different data outputs. It could be

augmented data, it could be multiplied

data, it could be, you know, all sorts

of things.

>> Doing this basically turns your data

flywheel from very limited to one eye

capture demonstration, that's what I

have to one to many. And we want to get

the one to many. So the higher fidelity

the simulation, the more complex

problems that we can start simulating,

which means that we can generate more

and more data.

>> Got it. That makes a lot of sense. So

sorry I kicked you off track a little

bit. So we're talking about pick and

place as well as spinal surgery. Uh walk

us through the next step.

>> Sure. So the first is data, capturing

data, uh generating data, augmenting

data. The second is training your model.

Now there's typically a few models that

might go in place for a robot. It's not,

you know, we're headed towards end to

end, but today it's it's kind of a

mixand match. So you'll have a

perception stack, something that is

classifying objects and poses. I want to

if I want to grab something, I want to

know what pose it is. So I know, you

know, how do I angle my my hand in order

to grab it? And then more more

importantly, where do you want me and

how do you want me to put that object?

So how I grab the object is also

influenced by how I need to place the

object. if I need to grab this and then

flip it upside down, maybe it's easier

to grab it upside down for the robot and

flip it versus doing this and and so

there's ways that um that you have to

think about what the robot is doing in

terms of its uh you know how it's it's

generating the trajectory. So you you

train a model for some of this. Maybe

you'll you'll train a skill inside a

simulation and then you can put them

together in a robot stack which would

have perception meaning I know how to

navigate around an environment. I can

see the environment. I know how to

perceive, you know, what's around me and

and identify obstacles, things like

that. I have a policy that once I get my

body, so imagine that a a skills policy

for manipulation, it's basically how do

I get my hands to where they need to be?

And so one part of the stack today is

how do you just move the body? The other

part of the stack is what do you do with

the hands? And so once you get your your

hands to the location, then it's okay, I

want to start doing a task. And so you

want to validate both of these things

inside of the robot stack before you

deploy it on robot. And so you can do

this in simulation. We call it software

in the loop testing.

>> Okay?

>> And so software in the loop testing is

where you simulate the robot and the

world. And then after that passes, we go

to this thing called hardware in the

loop testing, which is, you know, one

step before real deployment. Hardware in

the loop testing is where you simulate

the world, but you use the real hardware

component. So we use that third computer

and we feed the simulated data from the

second computer, the omniverse computer,

simulation computer into the onboard

edge computer. Oh,

>> and so the robot doesn't actually even

know that it's not out in the real

world.

>> It thinks it's doing spine surgery. It's

placing.

>> Yeah. And so we're going to feel and

then after that you can go into

deploying in the real world. And so this

is that end to end process data training

evaluation and then validation

deployment.

>> That's so interesting. So it's really a

matter of do you have like certain

specific buckets of skills that

determine what kind of tasks you can do?

like is it about building a skill

library to make a generalized robot or

is it like describe a little more about

how the capabilities for robots grow

over time?

>> Yeah, you you you um I think you you

framed it perfectly, the skills library.

So, we're going from specialist to

generalist. Think of a specialist as I

can do one very very specific thing very

very well.

>> Yeah. That I can do this millions of

times a day and I will not mess up. A

generalist needs to be robust to

changing environment circumstances, you

know, perturbations, things like that.

Yeah. And so if we capture all the data

from specialist, you can train a

generalist. And then the next step after

generalist is creating a generalist

specialist. So an equivalent would be a

child is a specialist in some ways,

right? They learn how to play with their

toys and they're really good at doing

some things, but they don't have enough

experience in the world to be able to

put all these skills together. Yeah.

>> I need to be posted for this. And so

then you go in and you get you get post

trained for that. And you get into the

real world. And once you're in the real

world, you still need to learn on the

job, right? And so a generalist is like

getting out of college. I'm I'm a a

fully functioning adult, but I'm not an

expert at anything. I'm just really good

at existing. I can I'm I can go into new

environments. I can learn new skills,

but I can learn. That's the important

part. I can learn a new skill. Right

now, we're at the point where we're

trying to train atomic skills. How to

grab things? Well, how are you

manipulating things? Very these are all

the same skills that a toddler or

three-year-old is trying to work on. And

then over time, you take these and you

start building them like Lego blocks

together. And so, the difference between

um you know, shaking a hand uh and you

know, maybe using a pool keel somewhat

similar in in action. And so there's all

these different actions that kind of

combine to create these composite skills

over time. And so we're we're doing

exactly what you're what you're talking

about. Think of skills library as one,

you could train a policy for it or two,

you want to just capture and generate as

much data for that skill.

>> Sure.

>> And then over time we can put them

together into these end policies which

is, you know, large multi-killed

policies.

>> Yeah, that that makes a ton of sense.

That's a that's a really exciting way to

think about it because a I think it

mirrors how like humans think about

their own learning process, their own

training process, and their own

validation process for lack of a better

word. Like how good am I at this skill,

right?

>> Um so validation is a question I'm

actually really interested in. Right? So

surgery versus pick and place. I need to

be much better with my hands to pass do

a successful surgery than to pick and

place a box. How do you know how do you

decide how good a robot is at a specific

skill? Sure. Not just that it can do it

but

>> yeah that's that's a good question. So

um one is is how do we evaluate? So for

instance we've released something uh

just recently called Isaac Lab Arena.

And so when you train a when you train a

skill in Isaac Lab which is our it's our

framework for robot learning. You want

to test that skill a b against a variety

of different environments right? So

imagine that um you know using

chopsticks.

>> Yeah that's a great

>> doesn't matter where you are. You're

going to use chop I could use chopsticks

to pick up pizza. I use it to you know

eat chips so I don't get my hands dirty.

Like you can use it's a skill. It's not

related to a specific food or a dish.

It's you can use it as a as a tool. And

so if you imagine I've taught something

to use a chopstick, but now I want to

have it do all sorts of objects, pick up

all sorts of different objects, you

know, like glass noodles and big pieces

of something and D. And so you want to

have all these environments laid out.

And so Isaac Lab Arena gives you the

ability to create all these different

environments and scenarios very easily

like Legos. And then you can create this

large library of scenarios to test

against. So that way as you're training

a policy you can see how it's performing

in not only one environment but in the

environments that you matter to you.

>> Yeah.

>> Um so that's that's one way. The second

is the hardware configuration really

matters.

>> Just because your policy can you know

you're likely able to train a policy

doesn't mean that the robot itself is

mechanically going to be able to do

whatever you need to do

>> and manipulate the chopsticks. what kind

of chopsticks

>> dexterity matters and that's why you

haven't seen many dextrous hands yet is

because a lot of the hands that we had

um they were lower than 22 degrees of

freedom um you know the human hand is is

pretty high up there you know even

between like you don't think about it

much but we use this palm area quite

often and most of the robots you see

they they grab and they they lose this

area in the palm they can't really use

it or rigid palm and and so

>> we're starting to see the mechatronics

become more advanced and start to mature

which means that the hardware can now

actually do what these policies should

be trained to do and so it's a mix of um

is your is your policy trained up enough

and you can say I don't have enough data

or I didn't train it correctly you know

things along those lines that's all

software bits and data bits on the other

side it's meatronically do I have the

right hardware in order to do it and

that's why companies like intuitive

surgical they have the hardware that can

do it and so now it becomes how do we

build policies around those hardware in

order to do certain things

>> and when you say policies you just mean

here are the not the rules but like the

general guardrails for how you should

manipulate or like

>> so a Um, a model like a perception model

uh would do things like classification

or pose estimation. So you feed in an

image and it goes okay this is what I'm

going to do. Um I'm just you you give it

an image and goes okay here's the pose

or you give an image here's

classification for a trained robot

policy. The reason why we call it a

policy is you know like what Pirates of

the Caribbean is like they're more like

guidelines.

>> The code is more what you'd call

guidelines than actual rules.

>> And so it's basically a set of

guidelines that say when you're in a

certain situation how would you react to

that situation? Right? So, I'm using

chopsticks and I've got this bowl full

of noodles. How am I going to approach

this versus if it's a bowl full of rice,

>> right? And so, that's the that's kind of

what it's it's trying to it's not a

black or white type of thing.

>> Yeah. And in surgery, you might really

care about that policy because how it

reacts to maybe something unexpected

during the surgery really matters,

right? So,

>> and functional safety, um the safety

boundaries are different between tasks

and and environments. So there are areas

where you have to be extremely safe. And

then there's areas where, you know, we

think of autonomous vehicles. If you

want to test an autonomous vehicle, you

want to have an autonomy autonomous

vehicle work. Um the safest is just to

build safety directly into it because it

has to be around human drivers. There's

no avoiding it. Um if I want a robot to

be safe today, I just move you to a

different room,

>> right? And so there's a there's a little

bit of u

>> the circumstance for that task. If it's

in a surgical room, then it needs to be

around humans. And so it has to be

extremely safe. It has to go through

plenty of certifications. If I'm just

trying to do a material movement, I'm

okay with it not being as safe. I just

make sure the environment itself is

safe. And so you don't have to build it

into the robot itself yet.

>> But eventually we will. It's just that

safety comes after fun, you know, after

the skills capability. Otherwise, what

what is it safe? The safest is just turn

it off.

>> Yeah. Right. And and I think that makes

that's actually a really interesting

like optimization problem too, right?

Because the safer you can make the

robot, the closer you can bring other

robots to it and still have it in

unison, you know,

>> and the more that you can put into a

single environment,

>> right? If you if you have a robot that's

only slightly safe in in some regards,

like it's safe in a very specific

setting, then you're limited by how many

you can put into any specific

environment, but if they are generally

safe, which is what our goal is

>> in end to end, you know, we we look at

humanoids because they're the hardest

problem. A humanoid has the problem of

locomotion, has the problem of dexterity

and manipulation, perception,

navigation, memory, balance, like uh and

then on top of that, because it has legs

and the upper body, it's this thing

called whole body control. Meaning that

you if you're grabbing a box, you bend

down and you pick it up, right? Like if

you're listening to the doctor, bend

down, pick it up, use your hips, right?

And that's actually whole body control.

Most robots aren't going to know that

out. They don't have this ability. They

only know either how to manipulate their

wheels and they can move around or then

they're doing it. That's why you don't

see very many robots that are walking

and drinking at the same time, right?

That's loco loco manipulation. And so

that's whole body control and we're

starting to get there. But humanoids

allow us to tackle these large problems.

That's why the ecosystem is is focusing

on this. If you can tackle the humanoid

problem, everything you build along the

way becomes plumbing and infrastructure

and tooling that you can then back

propagate into all of the industrial use

cases that are much more specialized or

narrowly scoped.

>> If you start with those, you actually

put yourself into into a corner. And

that's why we're we're tackling this

large problem. So you talked about so

tackling that problem means skills are

getting better over time. They're

linking together into bigger and bigger

skill sets so you can do more and more

right. One thing I'm really interested

in so when we talk about large language

models there are plenty of different

benchmarks for specific skills math

science literature are we going to see

equivalent robotic benchmarks?

>> Absolutely. The um the benchmarks that

you see today uh and this is exactly why

we built Isaac Lab Arena. So Isaac LLab

Arena is built on top of Isaac Lab and

it's basically uh an interface and a

framework for being able to design the

environment, the scenario, and the task,

right? Um I want to test a grasping task

inside of an industrial environment and

the scenario is that there's boxes that

are coming down a chute or something,

right? And so these three things are

like Lego blocks. If you could, you

know, manipulate these and create a

bunch of different scenarios from all

these existing Lego blocks, makes life a

lot easier. And so inside of the or

currently in the ecosystem there's

there's lots of benchmarks. Libro robo

bench behavior from you know Stanford.

These are all benchmarks that are

academic and used for testing the policy

themsel from an academic perspective.

One of the things that we're going to

start seeing more often is these

industrial benchmarks. I don't want to

just pick up a banana and place it on a

plate or you know any of those which are

absolutely necessary for the

state-of-the-art and and more frontier

testing. But once you get into

integration and I want to start using

this model to do things

>> I want to have my environment my

scenario my task and we're going to this

is where we're starting to to see the

environment or the ecosystem pick these

up start building their tasks so that

way they can test these these policies.

So you're going to see something very

very similar to math except maybe it'll

be micro assembly or maybe it'll be a

benchmark on um you know picking and

placing from a bin full of random

components and you have to do it in a

certain order or you have to do some

type of assembly task. So, you're going

to start seeing all sorts of these and

then we'll have categories of them and

it'll be very, very similar and it'll be

a whole library.

>> I'm really excited for that because

that's going to be very visually

engaging and I'm sure there will be

competitions around that whole

ecosystem.

>> And the cool part is whatever you do in

sim to some degree, you're going to need

to have it in real. And so, not only you

going to see uh these these scenarios

showing up in simulation to test, but

we're going to see the real world

equivalence of that because you need to

close the loop. So, you hear this often

in robotic close the loop.

>> Yeah. So you need to have a physical

space that you deploy these these things

onto these policies and these stacks

onto. You do the same task as you did in

simulation and you validate it in the

real world. And once you can validate in

the real world, we've now closed that

loop because now you've gone from uh

data and capture and all that all the

way to testing and validation. And now

this becomes that last bit of deployment

once we're like okay it works. What we

train does what we expected and now we

pull it. So you'll hear that quite often

and that's that's what makes physical AI

so hard is that we can't just leave the

validation in the data center in LLM

we're you know we have the privilege of

being able to test it in a data center

and leave it in a data center cuz its

whole life is going to be somewhat in a

data center and edge device

>> it's a SIM to SIM almost right

>> never really yeah exactly just like SIM

to SIM never really has to to actually

step into the real world and so that's I

think that's the the major challenge but

it's also the big fun of of this field

>> sure um does the loop ever go the other

way where it's like Hey, I have this

really difficult problem and I'm not

sure what kind of robot to even build to

solve it. So, I'm going to simulate that

problem first, see what kind of problem

like see the different solutions and

then build the robot that does the

solution the best or is

>> absolutely that you're asking all the

fun questions too. So the the what we're

seeing in aerospace for instance um they

can use simulation technologies and and

these AI agents that are able to modify

the design of thrusters and then

simulate you know the so similar to that

where they say we're trying to hit a

certain output right they want to have

some type of output for for their

engines and so it the agent is going to

keep going and optimizing and changing

until it gets to this until it gets to

this metric. something similar uh could

happen in robotics and this is something

that we're openly trying to uh trying to

research. Um first and foremost is the

embodiment the hand for instance the

hand the morphology of the hand. Do you

need three fingers? Do you need five

fingers? Is two fingers enough? Um what

exactly do we need to get the task done?

And it's so hard right now because um

like I said the the manipulator uh space

so like hands the ecosystem is is still

just is just beginning and we're

starting to see awesome hands coming to

market this year. And because of that, I

think what you're describing, being able

to look at the problem first and then

start evaluating, well, which robotic

components do I need in order to

accomplish this problem, it'll start

coming uh as as these uh you know, the

hardware matures because then you have

something that you can actually base it

against. You don't want to you actually

don't want to build a new robot all the

time, similar to like a car, right? You

you actually like having tier one

providers because it allows you to have

some consistency in your components.

Otherwise, if you have to manufacture

all of your own actuators and all of

your own internal components, it becomes

a huge drain on on on operational

resources.

>> That makes a lot of sense. So, even if

there is a hand that's better for a

specific task, you might want to default

to a generalized hand because then you

can just do a lot more with it besides

that one task.

>> And so, it it depends on the mix of

tasks. Uh so, an end toend robot, the

promise is that I could go to one work

cell here and then go to another work

cell here, dot totally different tasks

and use the same robot.

>> Yeah.

>> Right. That's the goal because that's

what a human can do. And so if we're

overly specialized, then we're still

stuck in specialists. So a generalist

allows you to go, you know, between the

field and that's that's where we want to

get to. And so understanding your mix of

tasks, the environment that helps us in

inform us on the hardware. Um yeah, I

think it'll go the direction that you're

describing, which would be super awesome

because there's going to be things like

assembly where okay, what would be the

best way to assemble this? Um and how do

you work backwards from there? I think

that that could definitely be a path

that we look forward to. Spencer, you've

clearly seen this whole industry evolve

very rapidly over the last couple years.

Is there something you're most excited

about or looking forward to? Like what's

the next thing that you're super pumped

about in

>> I am super excited for neural

simulation. You you'll hear this a lot.

Uh Cosmos is a world model and so it's a

neural simulator. It's been trained on

the dynamics of the world around it. um

these world models today uh are

improving and they're starting to show

extreme utility in as as part of our

policies. So when you look at Alpha Mayo

for autonomous vehicles that Jensen was

talking about, we can use these

reasoning models, these world models um

to actually be on board the car to

actually help us navigate through the

world. So the same thing could happen

for robots. There's in you know robotics

is um the oldest and newest industry in

the world in a lot of ways. The end

toend autonomy models are things where

the journey there is going to take us

quite a while and so we're going to see

quite a few evolutions of model

architectures between here and there.

Yeah.

>> And so we start with VLMs made a lot of

sense. Let's give robots the ability to

semantically understand their world and

reason about the world. The next step is

how do we make sure that the world

models are actually trained and

conditioned off of the types of inputs

that a robot would have. Meaning that as

a human I have perceptive input and

non-visual perceptive input. I have

contacts and I have action and all

these. So it's not trained on language.

We're not trained on language. We're

trained on all five senses. And so

language, visual, we need all the other

bits. And so as we start getting world

models that are able to um either be

conditioned off of these or output

these, we're going to start seeing an

influx of of totally new models and and

and capabilities. And so I'm super

excited for neural simulation for that.

Excited for neural simulation for data

generation and policy evaluation. It's

it's definitely a game changer for us.

I'm very excited.

>> What an exciting time to be alive.

>> Cosmos is going to be a it's an awesome

technology. I'm super excited. If you

guys haven't learned about it, you guys

go read about it. It's awesome.

>> I'm going to do just that. Thanks so

much for your time.

>> Wonderful meeting you, Alex. Thank you

for having me.

>> A huge thank you to Spencer Hang for

walking us through Nvidia's robotics

ecosystem, their three computer approach

to training, simulation, and inference,

and the biggest opportunities and

challenges in physical AI today. And if

you really want to understand robotics,

join me for NVIDIA GTC. You can register

for free with my links below and jump

into as many online sessions on robotics

and AI as you like. I'll announce the

winner of the RTX 5090 giveaway a few

days after the conference. So, make sure

to enter. Another huge thank you to

Nvidia for sponsoring my travel and my

media access to cover GTC Live and to

you for supporting the channel. Thanks

for watching and until next time, this

is Tickerol U. My name is Alex,

reminding you that the best investment

you can make is in you.