Loading video...
Video Failed to Load
To push self-driving into situations wilder than reality, we built a neural network world simulator that can create entirely synthetic worlds for the Tesla to drive in. Video below is fully generated & not a real video
1,700,083 views • 11 months ago •via X (Twitter)
36 Comments

We can synthesize new challenges (such as a pedestrian or a car cutting in)

We can make the AI drive in older challenges

You can also drive around in it like a video game

Same thing works for @Tesla_Optimus too

Can you simulate how it would've handled this situation?

We wrote an in-depth piece on the patents that make this tech so absolutely amazing. Check it out here:

What I love is there are startups whose sole product is real world generators that are half as good as this and this is only a feature in a back end system in Tesla.

One step closer To a halodeck

It looks so real!

World Model: Here we come!!!!!

This is cool, but also not good for $TSLA’s long term FSD moat. If simulation like this gets better and better, in theory a competitor without the real world data could simulate enough scenarios to fit all real world edge cases.

When is that game coming to PS5?

Wow!

Question! If this is the case, then why are you guys uploading gigabytes and gigabytes of data from our cars on Wi-Fi? 😅

The continuity between cameras is what really blows my mind. If anyone's interested, my review of Ashok's presentation on this is here:

Your “generated world” has basically no readable text… how exactly will FSD improve in a manner that has to adhering to road signs, when it isn’t trained well to read them in your generated training?

There can’t be simulated situations wilder than Bengaluru roads. Just being your cars to India and test them here. Once you master India, your cars can drive anywhere on the planet.

Is this procedurally generated in realtime, or are the environments pre-built and then can be explored / simulated? Imagine this being in a GTA game in the future, that'd be wild.

People aren't prepared for how common superhuman self driving cars will be by 2030.

That’s truly next gen.

These videos are in regular color etc. But doesn’t FSD now work on “pure photons”? If so how do the generated images convert to be usable in training?

When we will have taste of FSD14?

Still can't do this, can it? #DumbestStockFraudInHistory

Nice

Next level

Can you put my parking lot at my apartments into this simulator. I would love to see it learn how to avoid cutting through parking spaces straight into the same pole I park next to every day. The same pole I picks up on the sensors and has pulled out past every other summon?

This is mind-blowing 🤯🤩 Not selling $TSLA until I need to buy dentures 😅🤠

This is kinda crazy. I feel like my brain can't comprehend the fact that this is not real.

i'm not sure where you're coming from

That is 100% the right track! Great job guys!

Yeah we know. Sloppy work guys.

Explains why every FSD release is so bad until real people and real cities are in play.

Didn’t Elon said it’s stupid to have a simulator?

Had I not read the text I would have thought this was a real video!

The Matrix is real, Tesla is teaching everyone of our cars Kungfu!!

Impressive 😎
