Are AP1 and AP2 now using the same algorithms?

[trmmcd]

I currently have about 25K miles of AP1 AutoPilot usage. I have also closely followed the threads regarding AP2 Autopilot driving experience. Today I took my first highway drive using 2017.42 a88c8d5. It seemed like AP1 AutoPilot driving performance was smoother overall and it impressed me which is significant because I always thought AP1 was great.

However - I now started experiencing the phantom braking I read so much about in the AP2 cars. I also had my first ever "Take The Wheel Red Hands" experience on the highway while performing a lane change in light traffic. The fact that I am now seeing AP2 behaviors with AP1 makes me think that Tesla has now converged the AP1/AP2 algorithms for basic autopilot driving.

I think this is good news as it might lead to AP1 being able to leverage some simple AP2 enhancements and it probably simplifies Tesla's S/W maintenance. Also, if the smoother performance that I detected is now based on Tesla algorithms - it means they are really starting make progress.

Thoughts? Anyone else notice any changes with this release?

 

[Electroman]

That is one way to get parity. Wow, Tesla is innovative in ways I never imagined

I now started experiencing the phantom braking I read so much about in the AP2 cars

 

[DaddyP713]

Out of curiosity, does your AP1 car still recognise speed limit signs?

 

[mrElbe]

My AP1 car still recognizes speed limit signs. The new AP2 loaner I had recently did not. It used strictly map data. I tested this at a construction site where orange lower speed limit signs were present.

 

[sammyfan711]

I noticed the same thing last night(Slowing down for no reason, although it was on a city street rather than a highway) . It does seem like they changed the logic in autopilot one to our detriment.

 

[trmmcd]

The speed limit sign recognition is another good clue. I am pretty sure that behavior has changed. The limit used to change the moment I passed a sign. On unfamiliar roads today speed limit recognition seemed to be decoupled from the signs.

 

[avirtuos]

I highly doubt this. AP1 hardware was produced by mobileEye and is substantially different from AP 2.0. It is unlikely that the same machine learning models would run effectively in the AP 1 performance envelope.

 

[trmmcd]

I highly doubt this. AP1 hardware was produced by mobileEye and is substantially different from AP 2.0. It is unlikely that the same machine learning models would run effectively in the AP 1 performance envelope.

That is what I thought as well but I think the initial AP2 primarily only used Radar and one camera so it is possible. I also can’t think of another reason why AP1 would suddenly begin to exhibit AP2 behaviors.

If in fact this has been accomplished it is really significant. I wouldn’t expect Ap1 to gain much more functionality give the hw limitations but at least it hasn’t been orphaned entirely.

 

[avirtuos]

That is what I thought as well but I think the initial AP2 primarily only used Radar and one camera so it is possible. I also can’t think of another reason why AP1 would suddenly begin to exhibit AP2 behaviors.

If in fact this has been accomplished it is really significant. I wouldn’t expect Ap1 to gain much more functionality give the hw limitations but at least it hasn’t been orphaned entirely.

Its not about the sensors, the hardware they use to evaluate the models (GPU or otherwise) are extremely important as even when using the exact same software (models) and pre-recorded sensor input...simply changing which GPU type you run your model evaluation on can cause differing results. Some of these ML algos are impossible for a human to understand (unlike most software). This is why I doubt they are using the same software between AP1 and AP2...not to mention possible licensing issues with the MobileEye tech.

 

[ptsagcy]

My AP1 seems to be smoother as well, with the latest release. It definitely still reads speed limit signs instantly. I have noticed that it occasionally slows down momentarily when changing lanes, but not often.