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

FSD rewrite will go out on Oct 20 to limited beta

This site may earn commission on affiliate links.
I'm still not sure that the cameras are really enough to achieve reliable FSD.
That’s a concern of mine too, but it’s more the positioning.
About 6 months ago, I had come to a stop at a T intersection in a rural area. It was dangerous. My side vision was inhibited significantly by trees with growth at eye level, such that making a turn was risky guessing. Now think that the side cameras sit further back than our eyes, which means they were more ‘blind’ than me.
I wonder what peripheral vision overlap the front cameras have with the side cameras
 
Based on the videos I've seen of the FSD beta, the most concerning issue is the car's path at the end of turns. Many times on a leftward turn, the car tends to overturn and then overcorrects by turning right.

Other than this issue, the FSD beta is incredibly impressive, the turning path issue can be made more reliable over time by deploying the rewrite to the fleet and using neural network approaches to find the most likely path for a left turn into a particular intersection, using driver data. The labeling team can sort out deviations from the predicted turning paths.
 
The only contradiction I see is that high resolution radar is a crutch in exactly the same way Lidar is crutch.

Elon / Karpathy's whole point about lidar being a crutch is because it gets you to a local maximum quickly (i.e. gets you to demos asap). It has nothing to do with its resolution.

In terms of resolution, Elon and Karpathy are specifically against HD maps. They aren't against HD sensors. There's a difference here.
 

This video is pretty good since now we can really see the limitations. Interesting to see:
- It couldn't make either the first left or right turns
- It could not detect the lane markings properly to figure out it's a one-way street
- It missed a car in cross-traffic at a stop sign

I'm still not sure that the cameras are really enough to achieve reliable FSD.

This demo was pretty bad. There is barely any traffic and it failed so many times in 6mins. Is it me or is everyone over reacting to autopilot with a good turn once in a while?
 
  • Like
Reactions: jsmay311
He was confirming a statement that FSD isn’t possible with the cameras they have, without any facts. That’s common for this member here, and very different from questioning the unknown. :)
That is the GOOD part, he said in a later tweet that he has uploaded over 15GB worth of data from his Tesla and that it keeps sending more every time!

Good thing Tesla knows they have lot of stuff work in progress and collecting as much live data as possible. 15GB is massive to upload from the car...
 
The 15GB is being sent over LTE? A phone hotspot? How does he do metering?
Found the tweet.
upload_2020-10-23_18-4-52.png

https://twitter.com/brandonee916/status/1319670272039948288
 
The 15GB is being sent over LTE? A phone hotspot? How does he do metering?
https://twitter.com/brandonee916/status/1319670272039948288

He said it was traffic seen on his home network, and a lot of routers keep track of how much data is downloaded and uploaded.

I'm not sure if / how much data is uploaded via cellular from the car.

The little camera icon next to the profile selection supposedly sends an Autopilot debug report to the Early Access team. So Early Access users can manually trigger an AP snapshot to report bugs and issues with FSD/AP.
 
Andrej Karparthy, Senior Director of AI, took the stage and explained that the world is built for visual recognition. Lidar systems, he said, have a hard time deciphering between a plastic bag and a rubber tire. Large scale neural network training and visual recognition are necessary for Level 4 and Level 5 autonomy, he said.

I don't see how high res radar is any different in this case.

This is actually case where I would think Radar would be able to differentiate a plastic bag from a rubber tire based on the Radar return. The bag is far less dense, and wouldn't necessarily give as strong a return signal.
 
Why not? There are a lot of disengagements in these videos. (for now this is expected)
And many have nothing to do with construction.

FSD must handle construction, even full on blocked roads and re-route.

Sure a level 5 autonomous vehicle needs to handle construction, but is there any reason that FSD does? It would be great if it did, but don't you think it's more likely that Tesla would just say "Don't use FSD in construction zones" rather than have to worry about a whole new set of non-standard signage and edge cases?

Also on a separate topic, does anyone have any thoughts on how long it would take to go from full public release of the autosteer on city streets feature to anything above Level 2 driver assist? My personal guess is maybe some limited situational L3 (low to medium speeds in traffic, etc.), or perhaps a hands free highway L2 like Supercruise by late 2021. However, that's really not based on any deep analysis, just a gut feeling.
 
  • Funny
Reactions: jsmay311
Sure a level 5 autonomous vehicle needs to handle construction, but is there any reason that FSD does? It would be great if it did, but don't you think it's more likely that Tesla would just say "Don't use FSD in construction zones" rather than have to worry about a whole new set of non-standard signage and edge cases?

Here's an SAE question. Do vehicles need to be able to handle literally every conceivable domain before they can be deemed level 5? The second they can't handle X type of road they're level 4? Or would it be sufficient to say that if the car can travel from any arbitrary point A to any arbitrary point B in any conditions, it's level 5?

I could imagine Tesla determining FSD isn't great at handling certain road conditions, and just routes around them.
 
Sure a level 5 autonomous vehicle needs to handle construction, but is there any reason that FSD does? It would be great if it did, but don't you think it's more likely that Tesla would just say "Don't use FSD in construction zones" rather than have to worry about a whole new set of non-standard signage and edge cases?

Also on a separate topic, does anyone have any thoughts on how long it would take to go from full public release of the autosteer on city streets feature to anything above Level 2 driver assist? My personal guess is maybe some limited situational L3 (low to medium speeds in traffic, etc.), or perhaps a hands free highway L2 like Supercruise by late 2021. However, that's really not based on any deep analysis, just a gut feeling.



level 3 on the highway is probably first. There are a lot less variables despite the higher speed...
 
That’s a concern of mine too, but it’s more the positioning.
About 6 months ago, I had come to a stop at a T intersection in a rural area. It was dangerous. My side vision was inhibited significantly by trees with growth at eye level, such that making a turn was risky guessing. Now think that the side cameras sit further back than our eyes, which means they were more ‘blind’ than me.
I wonder what peripheral vision overlap the front cameras have with the side cameras
Peripheral vision at intersections is also a huge concern of mine.

The illustration below may provide some insight of the overlap.
onboard-sensors.jpg
 
  • Informative
Reactions: ord3r