New Elon Musk AI Podcast: Neuralink, AI, Autopilot, and the Pale Blue Dot

[EVNow]

He called perception and bringing objects into a vector space hard, which other like Mobileye call easy.
Holy crap, Talk about being behind.

Yes, I was surprised by this too. I'd have thought he would talk about figuring out how various actors would react to what the car does. Though, theoretically this falls in perception. I wonder whether he meant by vector space things like where other actors are and figuring out what they are going to do next.

Lex should have pressed him further on this question ...

 

[S4WRXTTCS]

Also, if NoA is safer than a human in stop and go traffic then why not get it approved as an SAE Level 3 system so people can watch Netflix? I feel like that would get a lot more people to buy Teslas and FSD.

My concern is they won't be able to get regulatory approval for even a limited speed L3 in stop, and go due to not having a proper hands free driver monitoring system.

To make sure the driver doesn't fall asleep while the car is doing the driving.

The rule with L3 is the driver has to be ready to take over, but most be given adequate time to take over. if the driver is sleeping then it screws the entire thing up.

So the car needs to be able to detect a person falling asleep to wake them up, and then put them into autopilot jail for falling asleep.

 

[heltok]

Yes, I was surprised by this too. I'd have thought he would talk about figuring out how various actors would react to what the car does. Though, theoretically this falls in perception. I wonder whether he meant by vector space things like where other actors are and figuring out what they are going to do next.

Lex should have pressed him further on this question ...

Teslas current FSD are based on the Multitask network with an RNN on top of it. This network has seen very many years worth of driving. It likely is extremly good at converting past 16seconds of video data into vector space at present and future time steps. There will be some uncertainty, ie it will predict that the space 10m in front of the car will be occupied with a probability of 0≥p≥1. If p is low enough it means drive with confidence, if p is too high proceed with caution. Over time proceed with less caution as Tesla feels confident enough with the system. Sometimes it will be wrong, but over time it will be less wrong as it is trained on more and more data where it has been wrong in the past. Operation Vacation is working its magic.

If you wanna see the vector space, check the autonomy day, at some point there the vector space was visualized.

 

[J1mbo]

Did I imagine it, or did Elon very briefly mention that they will use maps to know where to expect traffic lights? Is this his first confirmation of *cough* HD maps? Wasn't able to get timestamp unfortunately.

 

[Daniel in SD]

Did I imagine it, or did Elon very briefly mention that they will use maps to know where to expect traffic lights? Is this his first confirmation of *cough* HD maps? Wasn't able to get timestamp unfortunately.

Well to be fair he only said they're using maps and GPS to know where the stop light is. HD maps would use other landmarks to know exactly where to expect the stop light to be.
"You learn to drive in the parking lot, get things right at low speed and then the missing piece that we're working on is traffic lights and stop signs. Stop streets, I would say actually are also relatively easy because you kind of know where the stop street is, worse case you can geo-code it and then use visualization to see what the line is and stop at the line to eliminate the GPS error."
It sounds like they're well on their way to being where the industry was a decade ago!

 

[J1mbo]

Ah, see there I was thinking that geocoded features such as street furniture (which may be conditional) were in the "HD" part of HD maps, rather than just the map part ¯\_(ツ)_/¯

 

[diplomat33]

Did I imagine it, or did Elon very briefly mention that they will use maps to know where to expect traffic lights? Is this his first confirmation of *cough* HD maps? Wasn't able to get timestamp unfortunately.

No, not HD maps, just regular maps with certain features like traffic lights marked.

But I do think that Tesla will use Open Street Maps or something similar for navigating city streets since they are already using it for Smart Summon. Open Street Maps has traffic lights and stop signs marked on it. Tesla could pull that data.

 

[DanCar]

Waymo HD mapping uses multiple passes of lidar to create a 3D map with centimeter accuracy. Use of 2D maps is much different than HD maps. A presentation from Mobileye said there 2D maps takes kilobytes per mile versus megabytes for HD maps.