Autopilot differences between HW2.5 and FSD computer on residential streets

[Mardak]

I just picked up our Model 3 with retrofitted FSD computer, and I’ve noticed some slight changes both good and bad with seemingly the same software version 2020.4.1.

Here’s a residential curve that I would normally take at 20mph, and with HW2.5 set to 15mph would consistently fail while FSD computer consistently handles it at both 15mph and 20mph. It’s not a sharp turn like a normal 4-way intersection, and there’s only a lane line in the center of the road during the curve and curb on the right side.

Here’s the start of the curve to get a sense of how sharp it is:

And the end of the curve with line that ends and curb out of view:

Both HW2.5 and FSD computer display a confused right lane line:

I’m guessing the FSD computer happens to react just slightly faster to predict where a lane would go after the curve to not show the big red wheel “take over immediately” warning.


Here’s another curve that FSD computer now consistently completes; although it’s still a bit scary / late to react. This is probably difficult because the lane line disappears for an intersection at the very beginning of the curve, but perhaps FSD computer again is just slightly faster to pick up the line just after the intersection to immediately steer left.

Here’s a split from 1 lane to 2 where HW2.5 would consistently end up in the left turn lane, but surprisingly FSD computer knows to keep right even without a lead vehicle.

And for some regressions with FSD computer, here’s an intersection where HW2.5 would sometimes fail earlier in 2019 but later software including 2020.4.1 would correctly keep right through the intersection. FSD computer wanted to swerve left into the oncoming left turn lane and showed “take over immediately” because it knew it made a mistake of wanting to drive into the bus.

And this last regression is surprising as HW2.5 never had an issue while FSD computer shows “take over immediately” perhaps realizing the bike lane is not big enough for the car.

 

[Mardak]

One new behavior I noticed again today with FSD computer is automatic lane changes on city streets are able to complete where HW2.5 failed. Notably if a short intersection causes some lane lines to disappear, FSD computer doesn't immediately abort the lane change. There are plenty of instances where even with FSD computer the lane changes get aborted, so at least the ones that FSD likes are different from HW2.5.

 

[thewishmaster]

Interesting. I’m getting my upgrade next week so I’ll post some observations too

 

[Mardak]

Curious, have you noticed HW2.5 failing to complete ~20mph curves?

And for others that have had FSD computer for a while, is this curve-taking ability just "normal" behavior that you've always experienced?

 

[thewishmaster]

Curious, have you noticed HW2.5 failing to complete ~20mph curves?

And for others that have had FSD computer for a while, is this curve-taking ability just "normal" behavior that you've always experienced?

I don't have too many such curves near me to test frequently (mostly undivided roads in my neighborhood, can't activate AP).

The last time I tried a 90 degree turn with a continuous yellow line it wasn't looking good and was definitely going to cross the line if I didn't intervene; can't recall if this was on 2020.4.1 or 2019.40.50.7. I'll try that turn again with HW3 once I have it.

 

[croman]

Curious, have you noticed HW2.5 failing to complete ~20mph curves?

And for others that have had FSD computer for a while, is this curve-taking ability just "normal" behavior that you've always experienced?

My HW2 can do that and more (for some roads). I'm not sure what is going on with your comparisons but so many variables are at play so it is tough to really compare a situation but your overall observations about how it handles road markings and processes the width of lanes seem to indicate performance differences that shouldn't exist. The IC display is a different matter and FSD computer seems necessary but my car reacts to garbage bins (it just doesn't display anything in the binnacle/IC).

I can take it through S curves since 2018.50.x (prior to that it would repeatedly cross double yellows at some point in the S curve). Prior to 2020.4.x it would slow down properly before entering the turn. Now it barrels into the turn and slows only after the apex (which is pointless).

 

[Mardak]

My HW2 can do that and more (for some roads).

Oh yeah to be clear, there are a bunch of these similar looking curves that HW2.5 was able to complete without issue, e.g., especially if there are lane lines on both sides or a narrower street keeping the curb in sight or just a slightly less curvy bend.

I also wasn't expecting HW2.5 -> FSD computer to change behavior as it seems like it's the same neural network powering both while getting the same camera input framerate, so that's why I'm guessing FSD computer happens to have some slight timing differences where maybe Autopilot gets the neural network output 10 milliseconds earlier to make a slightly different decision to continue instead of giving up.

 

[thewishmaster]

Drove around and any differences are pretty subtle. The main thing I noticed is that HW3 is better at visualizing turning cars - orientation is detected pretty well as the car in front of you turns (HW2.5 often would show the car facing straight). Behavior is also smoother for acceleration/braking/changing lanes.

 

[Mardak]

HW3 is better at visualizing turning cars - orientation is detected pretty well as the car in front of you turns

Curious, is that pretty consistent for you? Some thing like showing oncoming traffic seemed to be better for a while but not so much anymore even on the same software version. Odd that it changes over time.

 

[thewishmaster]

Curious, is that pretty consistent for you? Some thing like showing oncoming traffic seemed to be better for a while but not so much anymore even on the same software version. Odd that it changes over time.

Yeah the oncoming traffic visualization is always a bit of a gamble for me. I’m sure the car sees all of it consistently, but it gets filtered out in the visualization most of the time. I thought it was actually pretty decently visualized in the first v10 release on hw2.5.

I haven’t noticed many obvious errors with turning cars (at least ones not filtered out), but maybe the true test will be seeing whether cars pulling out of a parallel parking spot get rendered accurately (something that never really looked right on hw2.5).

 

[Mardak]

Has anyone else noticed more WiFi upload with FSD computer vs before retrofit? I wonder if Tesla stopped collecting video data from older hardware perhaps related to the slight differences reported here. It does seem practical to avoid collecting and (over)training the network if the older hardware's timing differences might be bad for the network running on the new hardware.

 

[Bimmerville]

I noticed a similar improvement under similar circumstances with HW3. There are also some merges and lane splits that HW3 seems to do better with.

 

[Mardak]

There are also some merges and lane splits that HW3 seems to do better with.

I've been thinking about why FSD computer might do better with the lane split at the traffic light, and I wonder if there's some interaction with the left turn arrow on the ground. Autopilot's current default action is do the most common behavior, which is generally going straight, so it makes some sense that it might want to prefer the lane going straight vs turning left. But it could just be some accidental preference to the right lane vs left lane in this case as well.