Roland Busch

What we did is we created this hole in the digital world.

So we used a digital part of,

with a digital arm and a digital software and a digital camera and maybe everything.

And we trained basically a robot over and over again on this grip in the box with our technology.

And then we switched it on and the grip, the hit rate was still not satisfying 70 odd level.

It's amazing that because we trained, I mean, hundreds of hours virtually.

Then we used NVIDIA technology with a photorealistic ray tracing of these pieces, photorealistic ray tracing, different lights, trained the model over again, hit rate was jumping up substantially.

So these little details of having a normal representation of a digital part and a

really photorealistic one, made the hit rate coming up substantially.

And this is where... Also, there's a reason why when you train robots now in the virtual world, it doesn't really work.

There's a reason why we have so many people who are standing there with some handles and training robots to do a job and train them over and over and over again.

Because this is a real training on the real stuff.

These little details...

make a difference between an industrial application and one which you cannot use.

So what I'm saying is training models on specific data, on valuable design data, operation data, time series data, brings them to the level which we need in order to deploy them.

Don't underestimate the amount of data we have got.

I mean, I talk about generations, generations of design data for controls, for trains, for switches and whatnot, number one.

Number two is we have, I mean, I don't know how many thousands of machines we are operating.

We have machine jobs, machining jobs.

But if I then go to my AI guys, they say, I mean, okay, you make it now available, everything from Siemens.