Have we hit parity on Autopilot 2?!

ng0

Top-Contributor
Joined
Apr 11, 2017
Messages
996
Location
San Diego
Tesla Owner
Model 3
Country
Country
#1
So I saw this article just now and it got me excited about the quality of Autopilot 2!

https://electrek.co/2018/03/16/tesla-autopilot-update-first-drive-videos/

According to the article, there's some claims that Autopilot 2 may have hit parity with Autopilot 1. For years, Autopilot 2 has been lagging behind, but I'm really happy to see that they've finally caught up!

Any thoughts? Personal experiences?
 

KarenRei

Top-Contributor
Joined
Jul 27, 2017
Messages
1,644
Location
Reykjavík
Country
Country
#2
The big news is that they've gotten the new neural net working. They were previously bottlenecked by computing capability limitations and had to rewrite the neural net core to be able to operate faster on the hardware.

My expectations now are that it won't be as long between major updates.
 

Dan Detweiler

Top-Contributor
Joined
Apr 8, 2016
Messages
1,573
Location
Canton, Georgia
Country
Country
#8
Has this update enabled any additional hardware? It is my understanding that autopilot has been running on one camera, the RADAR, and sensors since the split with Mobileye. Perhaps this improvement is at least partially due to enabling more of the 8 available cameras on the car?

Dan
 

ng0

Top-Contributor
Joined
Apr 11, 2017
Messages
996
Location
San Diego
Tesla Owner
Model 3
Country
Country
#10
Has this update enabled any additional hardware? It is my understanding that autopilot has been running on one camera, the RADAR, and sensors since the split with Mobileye. Perhaps this improvement is at least partially due to enabling more of the 8 available cameras on the car?

Dan
I'm hearing some reports that the side cameras were enabled (unconfirmed as far as I know) but most importantly the neural net is what's powering the new features
 

garsh

Dis Member
Moderator
TOO Supporting Member
Joined
Apr 4, 2016
Messages
10,881
Location
Pittsburgh PA
Tesla Owner
Model 3
Country
Country
#13
I just watched this video:


It seems obvious to me that his testing has one huge flaw. When he covers up the rear camera, take a look at the rear camera view shown on the LCD.

THE TAPE IS PARTIALLY TRANSLUCENT!!!!!! The cameras can still see *something* with that tape on them (1m20s).

There aren't enough :eyerolls: to convey what I'm feeling towards this video.
 

Dan Detweiler

Top-Contributor
Joined
Apr 8, 2016
Messages
1,573
Location
Canton, Georgia
Country
Country
#14
I just watched this video:


It seems obvious to me that his testing has one huge flaw. When he covers up the rear camera, take a look at the rear camera view shown on the LCD.

THE TAPE IS PARTIALLY TRANSLUCENT!!!!!! The cameras can still see *something* with that tape on them (1m20s).

There aren't enough :eyerolls: to convey what I'm feeling towards this video.
Funny, I kind of find myself asking "why" when I see these types of videos. I really don't care HOW it works nearly as much as IF it works. That's just me being my normal non curious self I guess. I bow to the "inquiring minds want to know" crowd. LOL!

Dan
 

KarenRei

Top-Contributor
Joined
Jul 27, 2017
Messages
1,644
Location
Reykjavík
Country
Country
#15
One thing about neural nets (and genetic algorithms as well) is that they love to "cheat"; if they can find any loophole to exploit to accomplish their task (for example, translucent tape), they'll take it ;)

One of my favourites is relatively recent; they had been training a combination neural net+genetic algorithm to play the old Atari game Q*Bert. But the neural net found a bug in the game which - for seemingly inexplicable reasons - allows it to stay on level 1 after it's won the level and keep racking up points. It basically broke Q*Bert in a way that no human has discovered for the several decades of the game's existence.
 

garsh

Dis Member
Moderator
TOO Supporting Member
Joined
Apr 4, 2016
Messages
10,881
Location
Pittsburgh PA
Tesla Owner
Model 3
Country
Country
#16
One thing about neural nets (and genetic algorithms as well) is that they love to "cheat"; if they can find any loophole to exploit to accomplish their task (for example, translucent tape), they'll take it ;)
Think about training the neural net given all various inputs:
  • dirty cameras
  • nighttime
  • driving during dusk/twilight
  • driving on blacktop w. white lines
  • driving on light concrete w. black lines.
The neural net has probably learned to look for the "outlines", or "transitions" to determine where the lines are. It can't rely on the brightness or even the color of road lines. If it can still see the transition - even if it now appears to be a difference between "light pink" and "slightly lighter pink" due to the tape - it can still make use of it. So this didn't "block" the cameras. It just made it a little harder for the cameras to "see".