EEVblog 1504 - The COOL thing you MISSED at Tesla 2022 AI Day

[EEVblog]

A very cool bit of electronics failure mode analysis that you missed at the 2022 Tesla AI Day presentation. And it's got nothing to do with the Optimus Tesla Bot!

Shock and vibration analysis of the MEMS oscillators used by Tesla:
https://www.sitime.com/api/gated/AN10032-Shock-Vibration-Comparison-MEMS-and-Quartz-Oscillators.pdf
MEMS oscillator for automtive applications: https://www.sitime.com/api/gated/SiTime-MEMS-Oscillators-for-Automotive-Applications.pdf

[Icarus]

I don't buy it. All of this in 9 months ! 4 years more likely.

[EEVblog]

I don't buy it. All of this in 9 months ! 4 years more likely.

One of the design engineers said 6-8 months at the end of his talk.
Of course it depends on the talent they hired and their previous experience. it's not like you'd hire Joe Blogs engineer with zero robotics experience, regardless of how good they are at other stuff.

[nctnico]

I don't buy it. All of this in 9 months ! 4 years more likely.

One of the design engineers said 6-8 months at the end of his talk.
Of course it depends on the talent they hired and their previous experience. it's not like you'd hire Joe Blogs engineer with zero robotics experience, regardless of how good they are at other stuff.

Nahh, they likely just put some existing Python code together on a Rpi and called it a day after needing 8 months to get the Rpi going 
One of my sons is studying Bachelor level software engineering. He just started a few weeks ago and together with his buddies he made a facial recognition system (based on an Rpi) to check who is present in a class room based on a list with pictures. This sounds like a daunting task but they pulled it off within 2 weeks simply by using existing Python code.

[Psi]

6-8 months is totally achievable with A+ level engineers who have done walking robots before at other companies.
It wouldn't surprise me at all if some of the Boston Dynamics engineers are working at Tesla now.


With regard to Tesla Self Drive training data. As far as I'm aware, every Tesla cars is collecting driving data and video that is available to Tesla over the cars 4G connection. It's not something you can opt out of, afaik.
It's obviously not streaming all video back 24/7 or anything, but if Tesla are looking for particular examples of cars doing X maneuver at an intersection they can request that from all cars and then watch all the data and video flood in as every Telsa car finds examples of it.

The cars also do things like have the full self drive system active even when you don't have it controlling the car.
It looks for any serious deviations between the AI vs the human in control. If something happens and the human drivers does X but the AI would have done Y if it was in control then it gets flagged for investigation by Tesla to check who was correct. If the AI was incorrect the images and data get fed back into the AI training for the next version to fix the bad AI decision.

They also can run two versions of the AI at once on the car, the released version in control of the car and a release candidate in shadow mode.
When they have a new release candidate they can push it out to all cars and see how it would behave vs the released version in real driving.
Which helps them iron out bugs in new AI versions in a safe way.

Note: some of that info above is from old Tesla AI videos, i'm not sure how accurate it is for modern Tesla's, but its probably pretty similar.

[Icarus]

BUSTED !