Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

This site may earn commission on affiliate links.

Ashok Elluswamy (head of FSD/autopilot) keynote at CVPR'22 WAD on FSD​


Well worth watching.

Introduction
Classical -> Dense Depth -> Occupancy Networks for finding drivable space
Occupancy networks have some nice properties,
Occupancy network Architecture.
Static/moving objects, False Pedestrians - dummies, etc. -> every object is a moving object if enough force is applied!
Predict 4D occupancy flow - persistent.
Reasoning about occulsion.
Occupancy networks <-> Neural Radiance Fields (NeRF)
Full NeRF too computationally intensive to run in car, run cut down version for sanity checking.
Autopilot saving from crashes due to missaplication of accelerator.
Working on comprehensive crash avoidance. Search space large, take minutes to do full search, cut down version in car.
Come and join us.

I love the heading of this article “Uncrashable Car”. ❤️

 
I love the heading of this article “Uncrashable Car”. ❤️

It's not too unlike work done by stanford a very long time ago. Then it was avoiding traffic cones and they made humans drive drive through them at high speed. The car would pass and the human feel proud, but then replay would show that the human would have crashed into the cone and the adas saved the driver by adding minimal amount of extra steering to prevent the driver from crashing.

Here Tesla have extended this from the simple well defined cone example to a general solution. This matters.

Basically this can allow human drivers to drive without allowing them to crash. And it might guarantee that FSD will never crash, as long as the perception network and the prediction networks are accurate. So maybe push the system to be 10x safer than humans.


The talk was nice. First he presented occupancy grids from neural network and added flow to that. Then he showed NERF to generate 3D scenes from multiple camera poses, ie how to generate data for autolabel so it can generate labels to be consumed for training the neural networks in the car. Basically just an upgrade to autolabel.

Then he showed how neural networks can analyze a scene to see if a collision will happen, so that they planner can search through all possible outcomes and for each possible action see if a collision is likely. So the planner can decide where it can drive safely, comfortably and fast. This will be run through an optimization showed in the AI day and the best one will be selected. Very cool, will probably save many lives!
 
didn’t the P85D with introduction of the extra motor have improved efficiency?
I believe it was the S85D aka Model S AWD with 85 kwh pack.



"The AWD version adds highway range: The Model S '85D' version with an 85-kilowatt-hour battery pack has 295 miles of range (at 65 mph), the 'P85D' has 275 miles"
 
I’ve found the easiest - and most painful - way to stay humble is to focus on all the failures you’ve had as proselytizers either for purchasing a Tesla or for buying TSLA.

Successes in either of those goals are terrific, but why-oh-why have I been unable to convince that friend or relative to think clearly and do as we have done?

It is very, very humbling.
If I had a nickel for every time I screwed up, made a jackass of myself, pissed away money, got conned or got completely cleaned out in the stock market buying and holding loser stocks, jumping into the latest hot spike that deflates immediately or getting killed in options I would truly be a rich person.

Instead I just gotta deal with what Tesla has provided. So it goes.
 
I was hoping to get the FSD update but no luck, I've been on FSD Beta for about 2 months now and I love it. I understand only 1000 people got it so I didn't have high hopes this morning. From what I understand, after 1 week the next release will be to 10k people and then a week after that, all the FSD Beta testers will get the upgrade. I'm sure this timeline is dependent upon Tesla encountering few problems with the rollout.

It sounds like a nice strategy but I feel like a kid at Christmas and it's TWO weeks away!!!!

I'm hoping someone on this forum got the update and could let us know how it is.

Cheers to the longs
 
On underpriced, maybe. Ford said the Mach-E was no longer profitable due to cost of cost of components.

On the price raise of $8500, the timing is awfully fishy. After the IRA passed in the senate, ford and GM pretty much immediately raised prices for their EVs, but not their other ICE cars, as far as I know.
Ford Raised their prices about 2 days after Musk suggested Cybertruck prices would be going up. I suspect that has more to do with their pricing then the IRA bill. Particularly since their new pricing pushes many of their most popular configurations OUT of the price range for the incentive. They pushed all of their 320 mile range trucks over $80,000. So there is no incentive money to capture.
 

Ashok Elluswamy (head of FSD/autopilot) keynote at CVPR'22 WAD on FSD​


Well worth watching.

Introduction
Classical -> Dense Depth -> Occupancy Networks for finding drivable space
Occupancy networks have some nice properties,
Occupancy network Architecture.
Static/moving objects, False Pedestrians - dummies, etc. -> every object is a moving object if enough force is applied!
Predict 4D occupancy flow - persistent.
Reasoning about occulsion.
Occupancy networks <-> Neural Radiance Fields (NeRF)
Full NeRF too computationally intensive to run in car, run cut down version for sanity checking.
Autopilot saving from crashes due to missaplication of accelerator.
Working on comprehensive crash avoidance. Search space large, take minutes to do full search, cut down version in car.
Come and join us.
This is too long and detailed for many.... Some of it we've seen before. I would suggest everyone at least watch the last half of the last segment starting at 29:15. It's brand new and demonstrates how far Tesla Vision is being developed to potentially avoid all accidents. Wow! It's worth the 5 minutes it will take you....
 
That's what happens when you build a charging network because you have to, not because you want to.
I honestly don't know how they can be profitable when they are so unreliable. The amount of money they pocket is chump change compared to their operational cost. Seems to be an engineering problem and not a negligence of maintenance. Honestly how many people use these chargers before they break? 5 Charges? 20 charges?
 
Reports of 2022.16.3.10 rolling out:
1661058532531.png
 
I honestly don't know how they can be profitable when they are so unreliable. The amount of money they pocket is chump change compared to their operational cost. Seems to be an engineering problem and not a negligence of maintenance. Honestly how many people use these chargers before they break? 5 Charges? 20 charges?

It seems like they built it because they had to from the diesel cheating fiasco, and now that it is built, they don't give a s**t about it anymore
 
I honestly don't know how they can be profitable when they are so unreliable. The amount of money they pocket is chump change compared to their operational cost. Seems to be an engineering problem and not a negligence of maintenance. Honestly how many people use these chargers before they break? 5 Charges? 20 charges?
EA buys chargers from 4 different vendors so they likely have to deal with different support networks and parts for the various charging units. I haven’t seen the ones in the above tweet for example. Tesla’s in-house expertise here is a big advantage for reliability. When EAs chargers have reliability problems, they can’t just fix the problem at the source.