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

EAP uses different NN to Autopilot?

I used to have a 2017 P100D Model S with "Full Self Driving" - BAHAH! I struggled with AP2 during its infancy with phantom braking and all the usual complaints. Usually when passing a car or truck, it would suddenly slam the brakes on which is very disconcerting.

Needless to say, when I changed to Model 3, I opted not to go for anything other than "regular" autopilot. I was massively surprised at the difference of AP on Model 3 - basically the "dumb" version performed flawlessly.

Last night, I got tempted by auto-lane change, so upgraded to EAP... then went for a 100 mile drive that I do regularly on AP with no troubles. Lo and behold, I experienced some major phantom braking.

Does the upgrade switch over the NN and logic used - I'm not using NoA here, just plain ol' AP - but it seems the mere act of upgrading to EAP has forced me back into phantom braking territory.

Any thoughts/facts/evidence? I'm about to ask for a refund, since I can't be bothered risking safety and comfort through braking incidents when before the upgrade things were just fine. I'd rather have smooth AP than auto-lane change.
Same neural network, but different logic that uses inputs from NN. Would be a shame to get a refund a few weeks away from an upgrade that fixes phantom braking. Not saying that is going to happen, just a shame if it does.
What happens if you disable EAP? Does phantom braking go away? There is no phantom braking if you are trailing / tracking a car. Are you sure that isn't the difference?
I would turn off NoA (if on), and see what happens. I personally have to turn off NoA in places because it gets confused, and thinks I should be taking an exit so it slows down to get over.

If turning off NoA doesn't fix it I'd ask for a refund.

In any case it would make for a compelling test. Where someone tested an identical HW3 Model 3 with EAP (or FSD) against a HW3 Model 3 with only basic AP. Where the test was at least 500 miles of various highways/freeways.
I think the issue happens because EAP enables cut-in detection... speculation, but I remember there being chatter about this at the time.

Basically, my debugging instincts are telling me: it looks for indicators and a gentle move to the right (in the UK) and tries to predict if someone is going to cut-in to your path. Since lanes are narrow in the UK and lane discipline is *not a thing* here, AP just can't handle it so you get a lot of braking if you're passing someone who's wandering slightly to the right.

It's worse for lorries - in some light conditions, the orange tint to the lights on the back of their cabs can look like it's blinking as you pass (because it gets obscured and revealed a few times by the metal work of the trailers). To be fair, it does look a bit like they're indicating, but I can't have the brakes slam on every time I pass a lorry in the dark or in some sunny conditions that cause reflections.

NoA is definitely off... it really doesn't work in the UK (suggests dangerous/pointless lane changes all the time, requires manual confirmation anyway) - I remember all this from my Model S with NoA, and was very disappointed with it - it doesn't seem to have got better at all.

Even auto-lane change on dumb AP is pretty useless because of the 5-second window thing (i.e it aborts and swerves back if it hasn't completed the lane change within 5 seconds). I was hoping it would help me out on a few long-drives but... it doesn't.

Requesting refund...
  • Like
Reactions: DanCar
Yep - sorry - Auto Lane Change on EAP. I’m surprised that the simple addition of Auto Lane Change (which my hazy memory tells me the original MobileEye autopilot had back in the day - which is where I fell in love with autopilot originally) has such a negative impact on the experience/performance of “basic” autopilot.

Shame - in the UK I doubt we’ll see FSD for a very long time, and it seems like all dev progress has stopped on regular autopilot and NoA whilst they focus on FSD (understandably)