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

AP2/EAP starting to show signs of super human abilities

This site may earn commission on affiliate links.
I received 2017.44 firmware a week ago, just in time for a 1400 mile two-day drive from Phoenix AZ to Portland OR. I've been disappointed with AP2 performance (as many of us have) and the improvements I've seen in the past 11 months have been slow. OK, we've pretty much all covered that. **BUT** on this last drive I verified some almost "super human" abilities that AP2 now has. seriously!

First: lane following in curves with appropriate speed adjustment is now 90% good; occasionally on tight curves like I-5 north of Grants Pass (mountain pass, tight curves). Previous versions would radically slow down, forcing disconnects. This time it was quite good with only 10% of the overrides I previously saw.

Second: keeping pace with traffic is now quite smooth (usually with distance set to 2) at 70-85 mph. During the whole 1400 miles I had only two unexpected slowdowns. Both were from overpasses at the top of a substantial hill that fooled the radar. But it didn't last very long and I just accelerated to avoid screwing up the traffic behind me. Pretty good really.

Third: (this is where the super human senses kick in) driving in heavy rain, dark night, 75 mph along I-5 was PERFECT. I could hardly believe it. This was along I-5 between Eugene and Salem where there is no highway lighting, dark night, low clouds, fog, heavy rain and being splashed severely by trucks in adjacent lanes. I mean - I could BARELY see the adjacent lanes and was quite stressed about the trucks drenching me with water and temporarily blinding me.

However, to my amazement, the TACC and Steering kept on speed and within the lane just like the weather was clear and sunny. Wow! Of course I kept my hands on the wheel and hovering over the brake because it just seemed too good to be true.

I truly felt safer with the EAP driving than I would be driving myself. My night vision is good but not perfect and although the road is pretty straight, the radar seeing thru the splashing and heavy rain did a perfect job of keeping in the lane and adjusting speed to follow the (pretty much invisible) traffic ahead.

It would be cool to have some phone pictures showing the poor visibility outside the car but no way was I going to not be holding the wheel and peering intently around me. You'll have to accept my description of how little visibility I had.

One other thing: earlier in the day after a supercharge where it was raining quite heavily in Grants Pass, getting back on the freeway the TACC would not engage at first, message said something about sensor blocked. I thought crap, no AP until I stop and clean off the front radar. Well..... as I drive along at 75 manually I tried it again in about 5 miles and this time it engaged and stayed engaged. I think that the water build up on the radar cleared itself from wind after those 5 miles at speed even though it was raining. This turned out to be important to me as it gave me confidence (or at least an excuse) that the EAP would keep working during those heavy splashes I encountered. And indeed the system seemed to work very very well with no disconnect, slow downs, lane departure or scary behavior.

So I award AP2/EAP with this version my first **SUPERHUMAN** abilities award!
 
The camera sees the lane lines, not the radar. I had a similar experience driving in heavy rain and traffic in Seattle.
The lane lines remained blue the whole time even when I could not see more than 1 car length ahead/behind. Those cameras must be pretty darn good with night vision. Also I'm sure that I-5 is well mapped with great precision.

In modern avionics this is known as "sensor fusion" where smart blending of all available inputs and data produces the best output possible and the error probability score is constantly monitored to see if the system should disconnect if it "looses confidence". So it appears Tesla has really done their homework (as I would hope they would!!) on safety critical systems.
 
Cameras ssee the lines very cell in poor visibility, or with thin ice/snow cover, rain, even on my AP1
TACC is using radar, which is not significantly affected by any amount or rain, compared to the refltions of vehicles.
 
  • Like
Reactions: Gt1948
Thanks pilotSteve.

That's an encouraging experience on the plus side of the scatter graph.

Would be good to hear of many more positive AP experiences like yours before we can definitely say that AP is wearing its underpants on the outside :D
 
  • Like
Reactions: pilotSteve
Not to burst the OP’s bubble, but we’ve had that “seeing in the rain” characteristic since at least *.34 and quite probably before that.

It is the one use case in which AP2 is better than AP1.

I first saw it 6 weeks after trading in my AP1 car for an AP2 car, having no idea of how bad AP2 was by comparison. This was late May of this year - in blinding rain on a poorly-marked 2-lane state route northeast of Cheyenne, Wyoming, the car held the road and *saw* the road better than I could.

The next time this use case manifested itself was in mid-August of this year northbound toward Edmonton, Alberta from Calgary. Divided highway, with an almost complete absence of reflective paint in the rain at night. Despite that, AP2 did well.

But that’s it, folks. I’ve now had *.44 for a bit and there are still phantom braking events that increase the chances of my getting rear-ended, all kinds of constant disengagements in the same places that AP1 handled better, and truck lust of course is still alive and well. And there’s none of the website prattle about exiting the freeway automagically or passing a slower vehicle in front of me.

Less swervability is the main incremental improvement seen in *.44 so far.

Looking forward to the “fantastic improvements” claimed by senior management in the development releases, as soon as a) the fleet racks up the necessary miles to appease the regulators, and b) a decision is made as to when to retrofit current AP2 cars (owned by FSD investors) and with what.

You know - for features such as stop sign reaction as we saw in the December 2016 (AP2) video, first mentioned at the October 2014 (AP1) event.
 
Less swervability is the main incremental improvement seen in *.44 so far.


Are you comparing .44 to .42 or .34? In addition to much less swerving when lanes widen/narrow, .4x compared to .3x's other major improvements I've found are:

(1) Being able to make it through an intersection without diving one way
(2) Recognizes lane lines much more reliably, including recognizing some curbs as lane lines again
(3) TACC reacting to slow or stopped cars from a much longer distance than before — I would say it's nearly doubled. IMO AP1 still reacted further away, but this is a lot closer to AP1 now.

I just got .44 up from .42 from my uncorking appointment, and I still see some minor (not nearly as dramatic) improvements going from .42 to .44.


But at any rate, I think the coolest thing I've seen so far that makes me hopeful was @verygreen's discovery that the neural net can label "construction", and upwards/downwards slopes. This is the first time we've seen any evidence that the neural net was capable of more than just the functionality it was revealing to us (lane / car detection). Hopefully traffic lights / signs are also hidden in there or around the corner.
 
Are you comparing .44 to .42 or .34? In addition to much less swerving when lanes widen/narrow, .4x compared to .3x's other major improvements I've found are:

(1) Being able to make it through an intersection without diving one way
(2) Recognizes lane lines much more reliably, including recognizing some curbs as lane lines again
(3) TACC reacting to slow or stopped cars from a much longer distance than before — I would say it's nearly doubled. IMO AP1 still reacted further away, but this is a lot closer to AP1 now.

I just got .44 up from .42 from my uncorking appointment, and I still see some minor (not nearly as dramatic) improvements going from .42 to .44.


But at any rate, I think the coolest thing I've seen so far that makes me hopeful was @verygreen's discovery that the neural net can label "construction", and upwards/downwards slopes. This is the first time we've seen any evidence that the neural net was capable of more than just the functionality it was revealing to us (lane / car detection). Hopefully traffic lights / signs are also hidden in there or around the corner.

To *.34.

Am still experiencing your #1 above, although the intersection diving we reference may be different. On my end, consistently, for example on Sepulveda through Manhattan Beach, the car wants to become one with the center median past the intersection. Haven’t had diving the other direction that I can recall.

Agreed - was heartened to learn from verygreen that a new neural net is afoot. Having a bit of difficulty reconciling his actual findings versus senior management’s reports of much better performance - what nnets are senior management using in development *and with what data from what sources*, one wonders.

Maybe they’ll piggyback an enhanced firmware (stop signs) announcement onto the end of next week’s Semi event - although given that they just pushed the Chill/Easy Entry version with *gasp* actual release notes, the timing seems dicey.
 
That's great news about the car being able to see the lines in heavy rain.

I like the lines as an additional point of reference even when I'm not using AP. But, with AP1 the lines disappear with any significant rain. So it's good to see a significant improvement in this area with AP2.
 
  • Like
Reactions: pilotSteve
To *.34.

Am still experiencing your #1 above, although the intersection diving we reference may be different. On my end, consistently, for example on Sepulveda through Manhattan Beach, the car wants to become one with the center median past the intersection. Haven’t had diving the other direction that I can recall.


Don't get me wrong, it totally depends on the intersection whether or not it will succeed. I still need to override 1 in every 10-15 intersections, especially when I'm on an edge lane and there's some sort of boulevard/curb on one side, but before in .34 it was literally 90% failure. Even with 5 lanes, perfectly straight intersection, as soon as the line disappears my car happily nosedives for someone to my side. That has gotten dramatically better in .42. Not perfect by any means, but improved.
 
  • Helpful
Reactions: TaoJones
It is in CA... At 7 you'll have everyone cutting in front of you constantly in traffic... I run at 2 and go to 4 when it's wet but never over 4...
I am “in CA” and have always kept my setting at 7. It’s the only setting I’m comfortable with. Yes, sometimes people cut in front of me. So what? I still get where I’m going in good time. I don’t concern myself with thinking about how I might have arrived at my destination a minute or two earlier if I had my setting at 2.
 
The lane lines remained blue the whole time even when I could not see more than 1 car length ahead/behind. Those cameras must be pretty darn good with night vision. Also I'm sure that I-5 is well mapped with great precision.

In modern avionics this is known as "sensor fusion" where smart blending of all available inputs and data produces the best output possible and the error probability score is constantly monitored to see if the system should disconnect if it "looses confidence". So it appears Tesla has really done their homework (as I would hope they would!!) on safety critical systems.

The whole key to system integrity monitoring is that the error characteristics of every process are correctly modelled. For neural nets, there are many famous examples of nets misidentifying objects while outputting a high confidence / low error estimate in the result. I wonder how much we can trust the error estimate that is fed into the sensor blending filters.
 
Wanting to test the rain performance soon in CA. But with the recent updates, the lane following and ability to make it thru intersections (maybe because of looking out farther) has been great. Curves I could not do before are doable. I still do manual TACC speed adjust down on the tighest.

Great summary by our OP

old |v8.1(17.40.1 e29b97f)| AP2 |
new |v8.1(17.42 a88c8d5)| AP2 |
 
  • Informative
Reactions: GSP
Dec ‘16 90D AP2 22000 miles model S. today I was truly blown away by the ap system. Heavy rain, same experience as OP(must live in NW), and I went 30 miles on I90 where I could not have driven better. I couldn’t see the lines very well at times but I was rail straight. My car didn’t seem to have any trouble at all. Also a couple days ago I was “testing” the system in heavy crosswinds, and it held remarkably well. .42 and .44 are the most impressive I have seen since I have owned the car. It seems so much more confident now....
 
  • Informative
  • Like
Reactions: GSP and pilotSteve
Hey, I'm not your insurance agent nor family protector but is distance set at 2 really safe or appropriate for 70-85mph?

FWIW: at 70mph I'm at 7......
I didn't write clearly. In dry weather I usually use 2 otherwise I get cut in front of. However during the deluge I was at 4 or 5. I end up adjusting that distance setting quite a bit almost as second nature depending on traffic and weather. You are right: beware of close following especially when wet.
 
  • Like
Reactions: GSP
My only question is hydroplaning in heavy rain - the tires are selected for rolling ability - and traction is the complete opposite of fuel efficiency usually.

How was that during the heavy rain and trucks spraying you?
Well, a semi in the middle lane produces an inverted "V" shaped spray pattern, splashing BOTH lanes on either side with a huge amount of water and spray. Most trucks stayed in the middle (reasonable decision imo) but that meant cars are DRENCHED in water and spray as you passed the trucks.

Yes, hydroplaning is a serious concern which is why I was not "relaxed" about this hour long section of driving, but what can you do? So it was a considerable stress and workload reliever to me spending most of my attention peering ahead and trying to divine where the upcoming curves were and being spring loaded to take over if AP glitched. However in the end it was a stress reducer and made the drive much more enjoyable.