Tesla On Autopilot Slams Into Stalled Car

[Saghost]

I have LIDAR in my Audi A8. It is an option for 2900€ ($3,245). I doubt it would add $20,000 in price to a Tesla.


View attachment 412848 View attachment 412849

It's also very different from the LIDAR used by FSD candidates from other companies.

It's not obvious how capable a LIDAR is necessary (especially since it isn't obvious that a LIDAR is necessary,) but there's a wide range of capabilities and related price points in LIDAR, and folks claiming Tesla needs LIDAR mostly seem to be pushing for the top end pricy stuff used by Waymo and Uber.

 

[Triplett]

Negatory.

The fact that there are fewer Tesla’s on the road is irrelevant. The statistics are given per million miles driven. A mile driven in a Tesla is the same as a mile driven in an ice.

How can you support thot two? Where is your evidence to support that? Even if we assume it to be true, the statics also compare tesla drivers without auto pilot to Tesla drivers with autopilot, and Tesla drivers with autopilot go twice as far without getting into an accident as Tesla drivers without autopilot.

You also can’t prove thot three either. Most studies limit thier sample size to somewhere between 100 and 2000 samples (due to cost)... suffice it to say there are more than 2000 Tesla’s on the road that have driven sufficient miles to develope a sample size that provide a relatively high confidence level on the safety of the vehicles auto pilot system.

Again, wrong facts.

Thought = an idea in the brain.
Thot = that ho over there.

 

[5170304]

You're still missing the point. Tesla cars and owners are not representative of a sufficiently random and diverse group to be considered a quality sample for a statistical study. It doesn't really matter the actual numbers or how many miles driven. There are simply too many biased variables to be able to show real Autopilot effectiveness. If you can't understand why that is true then you don't understand how scientific studies are designed nor statistical significance works.

The stats that your "facts" are based on is equivalent to a study that attempts to measure the average weight of all people in the world by selecting a sample of people that all live in the same town.

If you want to truly study AP safety and effectiveness it needs to be installed on non Tesla vehicles and used by a sufficiently diverse group of people spanning the entire socio-economic and geographic spectrum.

https://www.cars.com/articles/whats-the-most-crash-prone-car-1420692141385/
This article gives a little perspective on the claims that Tesla owners are safer because they are driving more expensive cars etc. Turns out that claim is not necessarily true either. The most accident prone cars tend to be expensive fast cars as opposed to the least accident prone cars being pickup trucks and lower priced cars.

 

[Glamisduner]

This is a common situation, but I thought AP was supposed to use the radar to see in front of the car in front of you?

I swear some SUV drivers do this crap on purpose sometimes. Speed up then dodge out of the way at the last moment knowing anyone following them will not be able to stop for the obstacle. It's not just for other cars but also for road debris etc. Negligence of the larger cars in front of you can cause accidents for cars following behind them. Granted, you should travel far enough away for you to stop in time, but that's no guarantee the person behind you will be able to stop as quickly.

 

[Saghost]

This is a common situation, but I thought AP was supposed to use the radar to see in front of the car in front of you?

I swear some SUV drivers do this crap on purpose sometimes. Speed up then dodge out of the way at the last moment knowing anyone following them will not be able to stop for the obstacle. It's not just for other cars but also for road debris etc. Negligence of the larger cars in front of you can cause accidents for cars following behind them. Granted, you should travel far enough away for you to stop in time, but that's no guarantee the person behind you will be able to stop as quickly.

It does. But the radar has a difficult time telling the stopped car in front of you from the overhead sign and the soda can pointed the wrong way.

Tesla addresses this with fusion with the camera's neural net, but it isn't perfect yet, especially on older systems like this car had.

And this case had a distinct shortage of reaction time available once the truck came into view.

 

[PoitNarf]

I generally agree with you. While I certainly don't know the ins and outs of how Tesla magic works, multi-layer neural networks nowhere near as simple as you've suggested. There aren't that many specs on their neural network at all, aren't Training them results in multitudes of weightings, feeding through layers of convolution and deconvolution, soft max normalizations, feedback loops and probably some voodoo too. Have a look at the diagrams on this post, and I think you'll see what I mean: Tesla leaks info about new self-driving computer in latest software update - Electrek

I was under the assumption that the NN is primarily for image identification and recognition of patterns. Does the NN also control driving behavior based upon the NN deciphered camera feeds? Or once the NN has done its job with analyzing the environment around the car does non-NN code then use the gathered environment data to make driving decisions?

 

[Saghost]

I was under the assumption that the NN is primarily for image identification and recognition of patterns. Does the NN also control driving behavior based upon the NN deciphered camera feeds? Or once the NN has done its job with analyzing the environment around the car does non-NN code then use the gathered environment data to make driving decisions?

Tesla doesn't believe in one neural net to rule them all.

Instead, they have a bunch of them doing different things even within the camera image analysis section.

I'm not sure I've ever seen it explicitly stated how they control the car. I think we know from the investor event and the various folks digging into code that Tesla has neural networks that take the camera data and turn it into an understanding of the environment, and then that understanding gets handed off to something else to provide the reactions.

That something else is probably another neural net, but I haven't seen that stated; it could be hard coded reactions.