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

Autopilot lane keeping still not available over 6 months after delivery

This site may earn commission on affiliate links.
Let us not forget that all AP-capable cars, regardless of if AP Convenience Features are enabled or not, come with safety features like Blind Spot Monitoring and Automatic Emergency Braking that uses the hardware. Once Tesla decided to include those safety features, it had the hardware available to leverage for the convenience features.
 
I think most of the cost is R+D, not actual parts. The sensors are relatively cheap, by which I mean we're not talking about $3k in parts. It's obviously cheap enough that they'd rather install them on every car than split up the assembly line in any way to parcel them out as needed.

Yep. The way I see it the extra cost is a "tax" to get people to buy into the program from version 1.0 instead of waiting until version 3 and then paying the same price as everyone else and getting to take advantage of all the R&D done over the last few years.

Fair or not I see why they are doing it.

You could look at it as though the system should cost $3,000... But you are getting it at a discount for $2,500 by agreeing to sign up early and wait / test / use early versions. It may eventually just be $3,000 ... who knows.
 
I'd love to see the source for this. That last line in particular is extremely interesting. I'm probably wrong, but to me it seems as if it's trying to say that the first use of 360 8-camera fusion will first be seen in a commercial vehicle this year, using 5 EyeQ3 processors. There are only 3 or 4 manufacturers that I know of that are capable of this, one of which being Tesla with the Model X. It would be an astounding feat if true because such technology isn't expected to exist commercially until 2018 at the earliest.

Semi-autonomous driving this year: Mobileyes CEO Ziv Aviram - YouTube
More confirmation of the 5 eyeq3 implementation. There is a chance that Elon might have lied or stretched the truth about the Model X having the same autopilot technology and capabilities as the Model S. He might have meant (in sort of an avoiding the question sort of way) that at minimum, the Model X can perform the Model S functions as a minimum. We'll see when we start seeing the deliveries. The 5 EyeQ3 set-up only makes sense for a luxury car... so really it's down to Volvo, Mercedes, Tesla Motors or Audi. I doubt BMW.

Love seeing that the Autopilot released today is about as good as one could expect... and only to get better.
 
So now that I've played with autopilot for a day, I can see why they didn't release it sooner. By that I mean software-state, rather than calendar. This feels very much like a Version 0.9 or Version 1.0 beta. Good enough for wide release with warnings and caveats, but not polished by any means. Definitely would not have been safe or wise to release any earlier versions though it would be interesting to know what the major milestones were over the last 12+ months they've been working on this.
 
So now that I've played with autopilot for a day, I can see why they didn't release it sooner. By that I mean software-state, rather than calendar. This feels very much like a Version 0.9 or Version 1.0 beta. Good enough for wide release with warnings and caveats, but not polished by any means. Definitely would not have been safe or wise to release any earlier versions though it would be interesting to know what the major milestones were over the last 12+ months they've been working on this.

Well for one thing I'm pretty sure the navigation provider was replaced on the Model S with this. They need high quality maps that can be updated regularly to be accurate. Which is something they did not have before. You'll note that the map is now very seamlessly integrated into the instrument cluster. Before it a rectangle that left it not feeling like it fight with their UI. My guess would be the navigation provider was limited to that rectangular behavior and there wasn't a good way to fix it until they changed.

I've noticed that the car drives significantly better on the Interstate rather than city streets. Even though you're going much slower and the car has far more time to decisions. Granted city streets have more issues with lane markings (e.g. interruptions for intersections). But even where the lane markings are good it doesn't do as good of a job. You might say that this happens because you're following more on the Interstate, but even without someone in front of me it still does really well on the Interstate. I think the issue here is that the interstates all have high quality maps.

So you end up with a lot of interlocking pieces here. You can't release autopilot without better mapping. But you can't release that without all the UI code being done. There sorts of interlocking dependencies often slow down software projects. I'm sure there are a lot of other issues involved here. But this is one that I see a lot of outward signs of.
 
Well for one thing I'm pretty sure the navigation provider was replaced on the Model S with this. They need high quality maps that can be updated regularly to be accurate. Which is something they did not have before. You'll note that the map is now very seamlessly integrated into the instrument cluster. Before it a rectangle that left it not feeling like it fight with their UI. My guess would be the navigation provider was limited to that rectangular behavior and there wasn't a good way to fix it until they changed.

I do hope that they changed their navigation provider because the prior one was a bit comical. That said, I don't think that the autopilot is integrated into the maps. I've seen no evidence that it uses anything but the camera and radar.
 
breser,

I too noticed a significant difference between side streets (where we are not supposed to use it) and the highway. The car ping pongs on side street lanes, especially when it is not tracking another car. It is almost like the PID loop values are not yet well developed or the time constant is too large (too slow on updates/corrections). The strange thing is that the car does its best work with the most congested traffic. Its rock solid in those cases which seems to be where the feature has been targeted. That is good for people with crappy highway commutes like California.
 
breser,

I too noticed a significant difference between side streets (where we are not supposed to use it) and the highway. The car ping pongs on side street lanes, especially when it is not tracking another car. It is almost like the PID loop values are not yet well developed or the time constant is too large (too slow on updates/corrections). The strange thing is that the car does its best work with the most congested traffic. Its rock solid in those cases which seems to be where the feature has been targeted. That is good for people with crappy highway commutes like California.

In the notes it mentions that Autopilot uses both the car in front and the lane markings. If the car in front is not available it will use the lane markings exclusively meaning it could have trouble if the markings are not clearly legible. It is also more susceptible to measurement errors when trying to judge the position of the lines.
 
I do hope that they changed their navigation provider because the prior one was a bit comical. That said, I don't think that the autopilot is integrated into the maps. I've seen no evidence that it uses anything but the camera and radar.

Then you didn't listen to Elon's press conference where he specifically states that is is related.

Essentially the fine grained map data is the overide for corner cases where the system might otherwise ask for human input. If it has prior human input from ANY car on the network it will use that (data as a whole not single cases necessarily).


Adtionally it isn't just camera + radar. It is Camera + Radar + Sonar + Map data + GPS + steering wheel position + wheel speed sensors + accelerometer + anything else I missed.
 
Then you didn't listen to Elon's press conference where he specifically states that is is related.

Essentially the fine grained map data is the overide for corner cases where the system might otherwise ask for human input. If it has prior human input from ANY car on the network it will use that (data as a whole not single cases necessarily).

Did Elon specifically state the bolded? I would expect for a properly trained algorithm/map database you'd need several independent entries.
 
Did Elon specifically state the bolded? I would expect for a properly trained algorithm/map database you'd need several independent entries.

Those are my words not his, he discussed it with many more words and I condensed it.

I wouldn't wish the sound quality on anyone but if you want the content you'll have to listen to the call.

Tesla press conference for the Autopilot v7.0 software - YouTube

Unfortunately pieces of what he says regarding this are scattered out in multiple questions and answers and it takes hearing the entirety and combining it with other knowledge of the car to figure out what he means by parts of it.

But if you listen to the call you'll hear enough in Elon's own words that you should understand what I said.

And as a clarification "any car on the network" includes historical data of your own prior trips.

If you want to pay me a few thousand dollars I'll transcribe the call (it'd take me a day or two) and bold + annotate how it all comes together. Let me know if you want my paypal address :)
 
Those are my words not his, he discussed it with many more words and I condensed it.

I wouldn't wish the sound quality on anyone but if you want the content you'll have to listen to the call.

Tesla press conference for the Autopilot v7.0 software - YouTube

Unfortunately pieces of what he says regarding this are scattered out in multiple questions and answers and it takes hearing the entirety and combining it with other knowledge of the car to figure out what he means by parts of it.

Thanks. I've listened to parts of it, didn't get through the whole thing. I didn't get the same impression you did, but maybe I just didn't listen to enough of it.

From what I understood and would expect: A properly trained algorithm/map database would need several independent entries. Not just 1 person driving over the road and the database to be updated automatically. I could be wrong.