Good morning everyone.

Welcome to the last day of GTC. Now that

I know how to use this clicker, uh let's

get started.

Healthcare is fundamentally shaped by

physics. Right? So I want to anchor us

as I talk about all these innovations to

say this is this is not a net new

chapter. It's building on the continuum

that we've been on the for last two

decades at NVIDIA. But the industry

itself, right? A century ago with the

invention of X-ray, CT and ultrasound,

medicine got its superpower, the ability

to look inside the human body without

cutting it.

Then came image reconstruction, turning

raw signals into images, and accelerated

computing found its killer application.

Then we moved into imageguided therapies

where imaging started assisting in the

interventions and guiding the clinicians

and the surgeons. Deep learning had a

big bang and we moved from an age of

hardwaredefined rule-based sensor

processing to AI that can now start

learning physics. AI moved beyond the

scanner to ambient rooms where it could

understand, hear, listen, observe,

assist, and that's been the age of

agentic AI. And we're today at an

inflection point. Right? All of this

work leads us to the moment of physical

AI where AI can see, perceive, but act

in the physical world. And we'll see

that across embodiment. And it's all

built on the same foundation. We'll see

that across surgical robotics, some of

the hardest tests of physical AI. We'll

see it in clinical cobot embodiment,

care companions, supply chain cobots,

assistive humanoids.

And this is really needed for

healthcare. Think about it. We have

160,000 hospitals, 72,000 procedures, 8

million medical devices.

And while all of this data, all of this

sensor, all the devices can be

digitized,

care is actually delivered physically.

So if we imagine a world where there

will be all this assistance that can

act, the permutation and combination of

environments,

the O setups, the different devices, the

kind of procedure types warrant that

there needs to be investment in

simulation platforms that can allow for

these robots to learn in new

environments,

allow them to acquire new skills,

practice them before they can ever be

deployed in the real world. So I just

wanted to anchor us on the need in

healthcare for physical AI.

And why do we hear so much about

physical AI here at GTC is because there

have been three fundamental tectonic

shifts that are laying the foundation

for it. We in the healthcare team build

upon everything else at NVIDIA. we we

get the privilege of standing on the

shoulders of giants at Nvidia who are

building these foundations and build the

healthcare derivatives of it. So the

first pillar is world foundation models

right world models like cosmos can

understand and generate complex medical

environments that's what the team has

done with post-training cosmos for

healthcare surgical scenes and clinical

workflows

that's the first piece so you have world

models the second pillar is robotic

policy models like Groot and Groot H

that were announced at GTC that can now

enable these robots to have policies

that are generalizable across skills and

can translate from perception and

reasoning into action.

And the third pillar is the physics

simulation.

We have very high fidelity simulation

capabilities from classical simulation

that is fundamentally an accelerated

computing problem to now neural

simulation and generative physics. And

these proxy simulation engines are

coming very close to real world

performance. This is what's happening

across industries that's lending itself

to healthcare to have its physical AI

moment.

And that's why for physical AI

so far we've always spoken about AI and

the data challenge and the ability

whoever has the real world data mode.

Physical AI changes that because think

about it like in autonomous driving we

had all these numbers of hundreds and

thousands and millions of miles you had

to drive right but there is no way you

can drive through every terrain every

corner every edge case. So the way it is

actually there there has to be a data

flywheel investment to bring physical AI

to life. It has to start with real world

data because you have to ground these

models in reality in the physics of the

world we live in and you get these

pre-trained models but then there's a

continuous loop of post-training

and immense generation of synthetic data

and like I said you could be generating

this data with classical simulation

techniques which are well grounded in

the physics of the world these models

have to interact in but lately and more

so like my colleagues earlier shared

there are neural simul ation techniques

and reinforcement learning techniques.

But this is the flywheel that will allow

a data gathering strategy to have these

models in a continuous loop to learn. So

it's not just about access to AI ready

data. It's not just volume. It's also

that quality and coverage. And can you

you there's no surgeon, no hospital, no

one hospital would have seen all the

cases. So can you start leaning on these

simulation engines to generate synthetic

data to bring that diversity which

translates to robustness and

transparency for these AI models

and the first investment we had to make

was in this real world data. So a huge

shout out to the founding members of

open age who've been hard at it for the

last 18 months. Axel Kger, Dr. Nasir

Nawab, Sean Hoover, Madi Aizian that

started as a as a conversation has

turned into one of the largest release

of healthcare robotics data. Over 35

plus partners, the likes of CMR

surgical, moon surgical across 11

embodiment and 750 plus hours of

healthcare robotics data. This is the

first investment we had to make. And now

that all this data existed, we could

actually take the vision language action

models like the Grootbased models which

have essentially right there are two

component. There's a cosmos reasoning

engine which looks at the vision and

language and understand the environment

and intent to generate semantic tokens

that are then ingested by the diffusion

transformer with the robot states to

create actionbased policies or action

tokens.

And these models if you just take root

out of the box does not generalize well

for surgical robotics but open H made it

possible right so you have that

foundation and that's what we launched

at GTC this year group H is fully open

source it's available to get started on

GitHub to be download the weights can be

downloaded on hugging face so that's

great you have real world data you've

trained the action policy model that's

getting us much better performance than

zero short of any of these generalist

VAS. But it's not going to be enough.

And it's not going to be enough because

of the post-training synthetic data

generation loop that I shared about.

You're going to have to train this model

a lot more. And that's why the third

pillar of our investment has been in

Invidia Cosmos.

The world models, the world states in

which these robots have to interact,

they have to be as close as possible to

reality. So the first derivative we are

releasing for Cosmos uh H is for

transfer. It allows for controllable

synthetic data generation. You could try

uh you could start in Isaac same

omniverse to bring

real digital twin environments of the

human anatomy that's grounded in the

physics. So it's it's not generative AI

that can start hallucinating and then

you can be applying surgical uh transfer

to it to to create variations of that

world. So that is huge.

Surgical predict allows you to now start

generalizing and predicting different

action states. So it's many skills that

can be in one environment. Again, it's a

derivative of the foundational Cosmos

predict model from Nvidia and we've

essentially brought it all the surgical

data sets and the surgical domain

understanding so that it can actually

predict these different skills, practice

it in simulation and that simulation is

is very close to to to the real world.

And finally, the surgical simulator. Um

this is what uh the talk earlier

explained as well but this this is a

this is a foundation model of a kind

where we are now moving from rulebased

simulation to learnable simulation where

you can evaluate these policies and so

on and for the aesthetics of the slide

we've always kept it to benchtop but I

was hoping the audience here will will

would like to see some gory surgical

videos so I'll take that chance

Right from the left, what you're seeing

are segmentation depth masks bought in

from Isaac Sim and you apply Cosmos

transfer. In the right hand side is a

generative it's a generative AI powered

laparoscopic surgical video feed.

On the right hand side you see an image

and a prompt to say use the robotic

forcep to puncture a needle into the

soft tissue. And I have to admit I'm an

engineer. I have no surgical background,

but the video on the right does show

something looking like taking the

robotic foresip and puncturing into the

tissue.

And the action controlled the Cosmos

World Foundation model allows for you to

bring these embodiment and build

simulators

that can now start evaluating

different tasks across different

policies in different embodiment.

And that actually lays the foundation

for how we see

simulators for even some of the hardest

problems like surgical robotics being

solved

and all of these world foundation models

the data set we bring it and we anchor

it into our platform called Isaac for

healthcare. So I'm essentially making

sure you all know how is all of this

being released to developers

and Isaac for healthcare. In Nvidia

Isaac is the robotic platform. Isaac for

healthcare are domain specific

extensions on Isaac. So what do we go

do? We go find what are the domain

specific hard challenges. It includes

sensor simulation libraries, simulation

and synthetic data generation pipelines.

Some of them leveraging the foundation

models I just shared and models and

policies like Groot and others. We also

publish in it end to end workflows,

blueprints, developer recipes that

allows anybody to get kickarted

and it's a way of translating all the

technology foundations that Nvidia has

done in the three pillars that I spoke

about world models, action models and

simulation technologies and translate

that to use cases for the healthcare

industry.

So double clicking on some of these

domain specific elements, we have uh

sensor simulation libraries for

ultrasound, right? That's allowing you

to understand end toend simulation

pipelines from tissue acoustic

properties to probe configuration to be

able to generate real realistic scans

using NVIDIA's optics.

We also at this GTC have released X-ray

sensor simulation that allows you to

understand the end-to-end physics of

that pipeline from CT derived volumes to

ray propagation to detector response

and this allows you now to generate

synthetic sensor data at scale. So it's

a fantastic release. Isaac for

healthcare v0.5 sensor simulation

libraries available today. Uh in terms

of the synthetic data generation

pipelines

we publish these best practices. Some of

you might be using Isaac and others but

a lot of them are for industrial AMRs

and so on. And so we are really trying

to close the gap for the healthcare

developers. You can be building digital

twin

by bringing your own patients because

the world a lot of healthcare lives in

is the human anatomy and the human

physiology and so you can be leveraging

some of the models from manai to

generate a synthetic MRI data. It can be

grounded in a real world CT scan. You

generate a 3D mesh out of it. bring it

into Isaac's sim and from there you can

apply cosmos h and you have this engine

to be able to now understand the human

anatomy and create all sorts of uh world

states with that understanding

we have a pipeline to bring your digital

twins these are the physical

environments the robots have to operate

in and it's the same pipeline you you

create your digital twin you bring in

your digital twin assets you can use

neural reconstru instruction to multiply

that allow for those interactions and

those environments to be permutate

permutated and combined in Isaac's sim

cosmos transfer and you have the same

pipeline to essentially create the

foundation for this simulated world for

the robots to really learn to be robots.

We have the surgical video generator

that leverages the Cosmos H predict

model that we just shared and you can be

generating uh your synthetic data out of

that. And we also have a pipeline that

leverages the robotic generative physics

simulator. And all these pipelines are

available to you today to to just use

them out of the box and see how these

technologies connect together. But we

also publish all the post-training

recipes, the pre-training recipes, the

data sets we used for all these models.

So you could be in the room and

thinking, but this is all on

laparoscopic procedures and that's not

my domain. It's fine. You could you

could be taking all of that best

practices and adapting these models to

your use cases.

And this is something how the the

simplest view that we could get to for

the the physical AI robotics workflow,

right? So you start with that

pre-trained Groot H model that we spoke

about that's been trained with the open

data set we've released. So you start

with the pre-trained open Groot H model.

And now you need to multiply the data

and the environment variables and the

world variables Groot has seen. And so

you could do that in with real world

data. That's fantastic. You could be

doing that in classical simulation by

bringing those anatomical models from

your CT and MR scans into Isaac SIM and

applying Cosmos to multiply that or you

could be doing it with other generative

AI techniques. What you get out of that

is a post-trained Groot H model that now

understands has seen more variability.

But you still it still needs to practice

new skills. Maybe it is pre-trained on

suturing and you needed to get to

subtask autonomy of needle pickup. And

so you Isaac lab gives gives that

environment for that continuous

reinforcement learning imitation

learning to acquire these new skills.

And then Cosmos H simulator that

learnable simulator can actually become

the test bed for all these policies to

be evaluated.

And we shared in the paper that was

published uh by Sean and the team that

the policies like Grood Pi Zero and all

other VAS when evaluated in this Cosmos

based simulator

mimics the ranking of real world

performance.

That's really accelerating the

development time. Imagine having to

validate each policy on the physical

embodiment in the lab. That's days and

weeks. And here you have a simulator and

you can get that answer in a few hours

once it's at real-time performance even

faster. And then you have the simulation

engine and you know exactly why the

policy is not performing well and you

can go multiply more data. So this is

all about accelerating the researcher

and developer cycle.

Uh we've had the fantastic opportunity

to working with some of the pioneers in

the industry who are taking these

technologies seeing their value. Um and

it's likes of CMR surgical

uh J&J MedTech for their Monarch

platform has built on Isaac for

healthcare for their digital twin and

are investing with Cosmos based based

models for their data generation and AI

model training.

Moon Surgical, Medbot, Exar Labs, Lem

Surgical. Let me just highlight a few.

Like I said, CMR Surgical has really

pioneered that space. They were they

could see the value of the foundations

we are laying. They have come out and

contributed to the open age data set.

They are being uh leveraging cosmos age

the cosmos simulator groupage based

policy to start in their R&D space

around subtask automation for their

surgical procedures. They're also

leveraging the holoscan and IGX

platform.

Johnson and Johnson uh MedTech is using

Cosmos and Isaac SIM for their synthetic

data generation pipeline. And it's been

quite amazing in the last year. It's

just been 12 months. This is how fast

this field is moving to see world models

like Cosmos that understood roads and

signs and and our physical world, how

well they adapt and can be post-rained

to our anatomical and physiological

world. and our partners are seeing some

great success with that too.

Moon surgical is using Cosmos and Isaac

SIM and Isaac for healthcare uh for

system level autonomy. So they're still

in the environmental space but they are

able to import all the kinematics and

the robot states to really have that get

that system level autonomy for the robot

to get prepared before ever touching the

patient.

And so everything I shared so far was

around surgical robotics. We believe

that is the hardest test of physical AI

and that's why we wanted to apply our

technology foundations to it. But it is

the same foundations that go beyond

surgical robotics too. And that's why at

this GTC we've really expanded Isaac for

healthcare beyond clinical devices and

surgical robotics into hospital

automation. Project Rio was that. I hope

you've all had an opportunity to to see

the demo. But it's essentially a recipe,

a blueprint that developers can get

started today to say how do I assemble

my digital twin? How do I create these

assets? Uh we have fantastic partners,

the likes of like wheel, softs serve,

who are masters at doing that. But once

you've assembled that digital twin, you

start your data flywheel, the data

collection efforts, whether through

teleyop uh through capabilities like

mimic gen, cosmosbased models, but

you're essentially in that second bucket

multiplying the data, the variations

that the robot has seen and then finally

getting to policy training and the

testing in the loop until you're

confident that this this policy can be

deployed and I can take it to the real

world. it it's the same workflow. Uh

we've had the privilege Peritas AI has

shown this embraced this as we've built

it along and they're showing it at the

show floor. uh we have an amazing video

with them but they've fully embraced

from Isaac for healthcare cosmos age

this synthetic data generation to

accelerate their development of bringing

their uh Perry that's the name of their

humanoid the Perry robot to to the O and

they've been doing their experiments and

evaluations with Advent Health uh so

this is this is not sitting in a lab I

think all these customers we're sharing

about this is out of that and we're

really getting we're hoping and we We

absolutely believe 2026 is the year that

we're going to see the deployment

challenges and the age of deployment

coming uh for these for these robots.

And so, okay, you get the data set, you

can generate all this new data, you can

make the robots learn new skills, but

it's all in simulation.

But the rubber really hits the road once

you're in deployment.

And one big release we are having is

with Holliscan 4.0. O it is the ultimate

the ultimate deployment stack for

physical AI. I would say it's it's

physical AI native and there are three

key features that make it so. It now

supports EtherCAT that allows you to

have these control to directly the uh to

control the motor of the robot directly.

It allows interoperability with the

likes of Ross 2NG streamer. So you're

not it's it's not a completely new

stack. It's very interoperable with your

existing development stacks hopefully

and it allows for GPU resident graphs

that are fully accelerated on the GPU

with no interventions with the CPU

allowing for that speed of light lowest

latency in that deployment time. So

hollowcan 4.0 was released at GTC. Folks

in the room who are roboticists, who are

thinking of their next design platform,

please please go check out Hollowoscan.

And the ultimate combination is

hollowoscan running on an IGX Thor. This

is our industrial-grade edge AI platform

for realtime safety certified compute

for robotics devices, factory

automation.

IGX thaw really matters because it has

eight times more compute and I just want

you all to take that in. Everything that

I spoke earlier about these VLMs and

VLAS and Cosmos based models, all of

that AI is going to land on an edge

computer. And so edge device design

choices are being made now. And you want

to have your AI platforms and your

devices futureproofed for this age of

AI. It's no more in the age of CNN's

right. So IGX Thor and Holoscan 4.0 is a

fantastic platform which is absolutely

physically I native

and this essentially makes the same

point right what we're doing with

holoscan is you have this speed of light

latency and then sometimes you hear but

I don't need it to go that fast. That's

not the point. The point is how much

more can you do in that latency budget?

Now

you could do a simple CNN and a

detection model. If you're here at GTC,

you're seeing where the field is going.

And the models of the future are not

detection CNN's doing singular tasks.

They are the WLMs. They are the WAS. So

you want to future proof those

platforms. And Holcan is built for that.

All right. This is everything I shared

with you. This GTC has been about our

healthcare physical AI platform coming

out. We had we have released the open

age data set amazing foundation models

all the simulation best practices

recipes uh packaged in Isaac for

healthcare

but there's one critical piece and that

is that as you're thinking about in

robots how they have sensors for

perception of environment robots that

have to operate inside the human body

their perception is medical devices our

good old CT and MR and ultrasound sound

sensors. And so I just want to quickly

touch on some of the new models we're

releasing there as well. These are our

medical AI models, but they also come

together in service of physical AI

because a large part of domain specific

physical AI is inside the human body.

And so for Nvidia's medical imaging open

models all built on manai hopefully

you've heard of that. Um we have segment

family, we have a generate family,

reason and raw to insights. This GDC we

are releasing uh some key key updates to

that

introducing the Nvidia raw to insights

for ultra sound.

This is this is the first time you're

hearing about how we believe and where

we believe AI is headed. This is AI

based physics sensor processing. So

directly from the sensor to recon you

have physics grounded AI from the raw RF

signal. It can re in real time adapt uh

do patient adaptive imaging.

It's production ready and it was all

built uh in collaboration with Seaman's

research team with Altera and Nvidia's

IGX [snorts] and the HSB really allows

for that intervention of the raw RF data

to be able to train this model. So

you're the key takeaway is that you're

able to AI is getting to a point where

it can start understanding and deriving

insights from raw physics

and maybe there is a lot more insight in

that sensor processing that that that

gets lost by the time it's in the

imaging domain. So please take a note

GDC 2026 you heard the first time about

raw to insights and we'll be sharing

much more about that in the in the

coming year. We're also introducing raw

to insights for MRI. Uh this came out of

our challenge in Mikai with the CMRX

recon challenge. The foundation model

delivers consistent highquality

reconstruction across scanners,

protocols, and sites. It's shown its

robustness. It's number one on that

leaderboard. And the point again is that

for all the medical device developers,

researchers and as you guys are thinking

of where's innovation in that recon

pipeline, we really think this is the

moment just like in 2017

or 20056 where CUDA was discovered for

recon. This is the moment that AI is

really going to come to the raw physics

sensor domain.

Uh we're also introducing uh NV generate

for CT and MR. It's a state-of-the-art

3D diffusion model for CT and MR

scanners. Uh, and this one's amazed us

uh, a by its interest by all the

customers that we speak to, but also the

kind of use cases. So as an example the

uh when we have to bring anatomy into

digital twins there's a lot more CT data

out there and we generate allows you to

uh take a CT image and actually generate

a higher quality more resolution MR

image which is still grounded in the

anatomy of uh of the use case and all

these generating models are not meant

for any sort of clinical decision-m

that's not where we think the use cases

are. This is all in service of

diversifying the data sets as you train

these uh new models and modalities.

Uh Nvidia's medical imaging open models

are integrated uh into Hopper's AI

foundry. So if you've heard of Hopper,

they are here. They have an AI foundry

platform where they host their

foundation models like Cury and others

and Nvidia's open models are integrated

in that. It's a great place to get

started. I see Robert in the audience.

So in case you want to learn more about

Hopper. [snorts] Um and we're also

seeing integration with the likes of

Phillips where where they're taking all

these foundation models like segment

generate reason

these open foundations and post-training

it with their scanners their specific

scanners uh details and settings to to

really get the value uh for for their

use cases. So huge shout out to Phillips

as well.

Lastly, I shared a lot uh and then

there's just two places where you can

get started to to soak it all in to get

started. Uh it's the GitHub NVIDIA met

that's where we are hosting all the open

models.

They are all the source code is all

under Apache 2.0. We've released the

pre-training scripts, post-training

scripts. We're really wanting to move

the ecosystem forward. So we we are

sharing in real time as we are learning

that's up there. Hugging face is where

we are hosting all the weights. Openh is

also available on hugging face to

download today. And Isaac for healthcare

GitHub is where you can see how all

these tools and technologies and models

and data sets can actually be applied to

to post train for your use cases for

your applications whether in surgical

robotics um in clinical medical devices

uh or for hospital automation any kind

of embodiment.

With that thank you all for your time

today and I'll take questions.

>> [applause]

>> Thank you Pera for the talk. Um for

people who have questions, please line

up at the microphones that are here in

the aisle and we'll go first person over

here.

>> Okay. My question is about autonomous

surgery. When do you think this is going

to happen and how much governance hurdle

you think the world has to go through?

It will depends on the countries.

>> Excellent question.

Surgical autonomy we believe is the

absolute northstar in surgical robotics

for for it to come. So that's we believe

is is long way out. What we are seeing

is subtask automation, subtask autonomy

and I think our partners like Medbot if

you go to SRS and other conferences they

have very grand visions about it. even

at their talk at GTC um Medbot is one of

the pioneers who's taking some really

innovative leaps in terms of the

regulatory stuff. It does come down to

which countries are we talking about and

it really the the the word autonomy has

a has a large gamut and so it really

then comes down to what is the subtask

automation and where where those

guardrails lie.

>> Thanks for the talk. Um, with these

approaches where you're passing the raw

data to the models, do you have some

kind of a hypothesis about what might be

missed in the non raw approach or any

kind of early hints of that?

>> I would say we still in early research

to even say that as we pass the raw

data, can it even start understanding

the the acoustic behavior and the

behavior outside of the imaging domain?

the the domain is getting accelerated

sooner and I think we will be at a point

where we'll be able to show what's the

before and after. So like the seammen's

ultrasound uh data set or and the and

the model that we're sharing as as it

gets applied further like what is the

what was missed in the before approach

that now that this AI can understand all

the sensor physics it's able to augment

or or predict um I think that's still

TBD but I do think this would be one of

the the most interesting topics in some

of the academic conferences I highly

recommend you connect with Sean Hoover

uh who's sitting right there and he's

leading a lot of this work.

>> Okay, great. Thanks so much.

>> Sure.

>> Okay. Also related to this question uh

the model that you just show uh going

from the raw data to imaging domain uh

you have several different kind of uh

physical sensor right are they are based

on the same architecture I mean the

model itself or it's a different model

to do the different thing uh different

raw input data

>> currently what we are releasing in the

two model variants we have so for

ultrasound we are releasing a model

architecture which is um trained with

the open age RF data that was

contributed by semens.

>> We don't believe that the model

architecture will be changing based on

the physics sensor modality but as we

continue to build upon it we'll we'll

share more um the the raw to insights

for MRI is a completely different

architecture and actually solving quite

a significantly different problem. it's

actually getting into the more recon

approach like an AI based recon approach

um as compared to just a CUDA based

recon approach.

>> Okay, good.

>> Sure.

>> Hey, good morning. I have two questions

on the Cosmos age the data generation.

So number one so uh how long can can it

generate the data like for seconds for

minutes or hours? And the second

question is is the user able to

configure what type of data it can

generate. For example, can can you

generate a data for a successful task or

for a failed task?

>> Uh I followed your first question. Can

you repeat your second question?

>> Sure. Sorry. Uh the second question is

can the model generate a task uh uh the

data for a successful task or for a

failed task or should is a user able to

configure that?

>> Got it. Yeah. Okay. Okay, so for your

first question, how many hours of Cosmos

data can it generate like as much

compute as you can throw at it, right?

So it's not limited by it's only these

many seconds of clip or or so on so

forth. Um so that hopefully answers the

first part of your question and your

second part it can. So uh your second

part of the question is is it able to

only uh generate cases of successful

tasks? That's not uh the case. Indeed,

Cosmos has surprised us uh with the its

ability to learn physics

>> even just by the the videos and so on.

So, it it does need that constant

prompting and post-training loop, right?

So, it's not like you generate all the

data and you train Cosmos and you're

done. You generate data, you train

Cosmos, you prompt it for the corner

cases you're looking for and is it quite

getting it, not getting it? And you

might have to try different approaches

for it. But it absolutely has shown that

it can learn the physics and actually

generate cases that it hasn't seen. And

that's that's really the point of

Cosmos, right? It is meant for those

corner cases. And so as rich uh as a

input you can give to it. We've seen it

performs well the best. It performs well

in all these different variants, but it

really performs really good when it's

when you have uh the anchored digital

assets from CT andMR that I shared

earlier with you. Um that really grounds

it in the physics of the of the world

that it's trying to learn about.

>> Okay. Thank you.

>> Sure.

>> Hi, this is Suresh from United Health

Group. Um thank you for the talk and

it's exciting to see you know how much

physical AI is progressing. Uh obviously

I'm part of the division in in pharmacy.

Uh so I'm interested to leverage some of

the world models uh like you know

applicability and I definitely connect

with you and I also let know our

colleagues on the healthcare side like

know where you know we can leverage

that. So do you know who I can contact

or can I connect with you directly?

>> Yeah I mean you can connect with me. We

have amazing NVIDIA team here. So we can

maybe chat afterwards and I'll get you

connected with the team.

>> Okay. Sure.

>> Thank you.

>> Hi, thank you for the talk. Um I have

three questions. The first question is

um how do you like when you go from a

simulation to deployment? Something is

going to break. So in your experience,

what is the first thing that breaks and

that simulation is still not able to

capture completely?

What is the first thing that breaks? I

think sim to real is not a solved

problem, right? Uh we are also making

those endeavors and the first things

some of the first things that break are

the the the

even just the because your digital twins

are not physics consistent and accurate.

Um the it it's it's system level issues

that we end up seeing with the port

connectivity issues. So it's not a

seamless switch today. I think that's

what we're trying to work out with uh I

don't know if you attended the earlier

talk where Madi and Sean shared with the

hollow scan. So if you are building your

pipelines with hollow scan then you can

test that deployment in simulation and

if you take that same stack our goal is

to minimize that gap of sim to real

deployment.

too many it's very hard to say what is

the one thing that breaks but too many

things break and the reason they break

is that it's it's a completely different

stack right so there still has a lot of

bespoke work to be done as you translate

all that simulation learning um but I

would say that's much more on the on the

system and the network and the and and

the system connector side of the thing

>> thank you my second question is that if

you have two systems one system is

trained completely on simulation, tested

in phases and then deployed. And then

you have the second system which is more

um traditional and conventional and you

test it and test you build it and test

it in phases with real-time deployment.

At the end, which system still performs

better or what's the difference between

the completely simulated system versus

the tested in realtime system?

I don't think the conversation will be

so much about systems that are fully

trained in simulation and systems that

are trained fully with classical

approaches. So as an example like I mean

maybe we can anchor the problem a little

bit. So if you had to say uh

I believe today surgical robotics for

their simulators these companies pay

multi-million dollars for that very

accurate uh simulators right but we have

to start now adding the variables it's

mult it's it's a cost variable it's a

development variable it's potentially a

multi-year approach so just to create a

com using classical simulation and

computing technologies to create uh uh

accurate simulator for a procedure

much more costly many years of

deployment development uh and then in

deployment I'd say I considering that is

the only system that's used in practice

the fidelity of the system is good right

now if you come to what we shared for

instance with the cosmos 8 surgical

simulator

the development and the cost times are

reduced by a magnitude ude and uh we our

early work is showing that the the

behavior of the policy ranking is is

quite close to real world performance of

those policies. So we don't believe a

world where all the classical simulators

will get replaced by this. We believe

that the people and the pioneers who've

been investing the likes of surgical

sciences who are building these

classical simulators will will converge

and there will be an hybrid approach and

there will be some aspects of that tool

tissue interaction or the [laughter]

some aspects of that physics that you

will need classical techniques and there

will be some aspects of that accelerated

development where a close proxy from a

simulator is good enough. Right? But our

hope is that that significantly reduces

a the cost and the development time it

takes to build these highfidelity

surgical simulators. Uh which allows for

the technology to to grow faster and

expand faster into multiple procedures

and so on.

>> Got it. So a hybrid approach, right? I

notice.