#Fly By Wire Model (5 Sep 2025)

93 messages · Page 1 of 1 (latest)

ancient wyvern
#

cool

#

<@&1386399863293345912>

spark lichen
#

Installing

simple jetty
#

In a parking lot. Grabbing now.

magic tartan
ancient wyvern
# magic tartan

This one torq made a pr like 6 hours ago at almost midnight pacific time

#

wild

spark lichen
#

Hopefully fly by wire means experimental gonna be fast

magic tartan
errant gate
#

I have an 80 mile drive on the interstate today for work, which model should I test first? gWMv4 or Fly by Wire?

gaunt remnant
#

master being firehose model now in this comparison

ancient wyvern
fierce oar
#

Now I gotta listen to the song

pseudo basin
#

Has anyone tried this model? How it compared to gWM 4?

long roost
#

It was just made available less than an hour ago 😂

#

But imma try it in a bit

pseudo basin
#

Is it so good that someone doesn’t want to finish the test drive for an hour?😂

gaunt remnant
mellow falcon
#

I haven't even got to try Firehose yet 🥲

ancient wyvern
brave oriole
#

Vibes from a highway drive to work.

Low speed torque requests are less than firehose, but recentering still seems to be good.

Highway lat is good, on par with firehose.

This models seems to identify cars further out more reliably and match speed quicker. With a lead it sat below the set speed and slowly crept up on them going around 70mph.

Have not came up behind anyone that has been stopped but I think this model will probably be better at long than previous models, will be interested to see what everyone else experiences. Will run again on the way home before trying gWMv4

odd obsidian
#

Almost a wow on lat. aggressive curves. Occasional left biasy on some straight s for my taste. Not as planted as gwm 4

spark lichen
#

Just did 26 miles on this one. I agree with @brave oriole it seems to detect leads from a further distance, long felt real safe. I ran full experimental in city and speeds are still too inconsistent but I used it for all the city driving. Curves felt good it did go over the white line a bit on the inner side of the curves but I don’t think it went too deep into the neighboring lane.

odd obsidian
gaunt remnant
#

hmm I dunno if I'm using this wrong but this has been horrible for me

#

not going around parked cars much just wanting to go straight into them or if it does go around them it stays in the opposite lane rather than going back

#

turns (with turn desires) have been terrible too

ancient wyvern
#

It does not respect stopped cars

spark lichen
#

It went around parked cars for me but I was very low speed like under 10mph

stuck hazel
#

Less complex architecture

#

We're still not allowed to merge green diffs, until monday

glacial quartz
#

Right after downloading this model I was driving down a road and it acted like it was going to hit a parked car. Thought I'd switch when I got home, but as I was driving down the road to my house it did the best job I have seen from any model at driving around parked cars. It gave plenty of room (which every model I have tried gets a little closer than I would like) and it even stayed out a little wider as there was another parked car about 200 ft further and then after passing that last one it centered itself perfectly.

Not sure why it did so badly on its first attempt though. Will definitely keep testing. Though it doesn't have as much steering torque as other models which is a bummer.

spark lichen
glacial quartz
gaunt remnant
#

yeah steering torque seems to be lacking on this model for sure

half yacht
#

This model has a horrible left bias

tepid flame
# glacial quartz Can someone explain like I'm 5 what this means?

Comma is doing a sprint (2 weeks?) where they only merge changes that delete code. They said no new models until this is over. In this case Torq said that this model came about from deleting code from their training stack so that it counts towards the red (delete) code sprint.

#

I guess we'll find out whether whatever he deleted made this model better/worse or unchanged relative to the others

glacial quartz
tepid flame
glacial quartz
half yacht
spiral leaf
#

Definite left bias, let it stop at a light. It stopped a little late, about 1 ft over the line. Then it very slowly crept about 3 ft. and tried to take off. Oops. Lat was OK.

ancient wyvern
#

Long is weak on this model

#

Blows reds compared to gWM4

#

Experimental gets too close

#

Regression compared to gWM4, but gWM4 is a pr too so I guess I can’t compare the two

tepid flame
#

Firehose has been great (and now in master) so that's what I would compare it against

ancient wyvern
upbeat matrix
fierce oar
#

gWM4 isn't red_diff right?

#

only this one so far

half yacht
#

gWM4 > Firehose Model > FBW

fierce oar
half yacht
#

I’m not sure gWM4 is

#

FH and FBW are AFAIK

fierce oar
#

kk

tepid flame
#

oof yeah this one's lat feels more sloppy and lazy than firehose. It also feels like it's accelerating too much which made me feel super uneasy for the first time in exp long, like it was going to ram into the back of someone if I didn't start slowing down manually

sick fulcrum
#

I didn't read reviews before trying Fly by Wire on 2025 Ioniq 5. (Lat only)

Less smooth than Firehose. More torquey. Might be OK except it seems to have more of an uncomfortable right hug on highways. After more highway driving I find its lane keeping to be inconsistent. Mostly fine. Sometime with uncomfortable right hug. Sometimes hugging left. It's mostly fine but less comfortable than Firehose on highways.

I noticed on local unmarked streets it's not very good at dodging parked cars. Firehose was noticeably good at that.

odd obsidian
#

the highlight of this model is the aggressive curves more aggressive than spacelab3v2 it seems

brave oriole
odd obsidian
odd obsidian
#

They should rename fly by wire to LCT lead car tracer if you follow too close like one bar of distance, it will copy the lead car position so if the lead car ping-pongs you’re gonna ping-pong, if the lead car left bias or right bias, you’re gonna do the same and so forth and so on, so give this model another chance and have the lead follow in mind

warm creek
odd obsidian
#

yeah observe the lead car you gonna have fun with this model

#

If you wanna prevent it just do the 3 second rule follow

#

and stay in the slow lane

odd obsidian
tepid flame
#

So I'm actually going to do a 180 here and say this model is quite good! There are definitely still some issues with lane placement and it failed a cone test I did, but the upside was that it was the most "human" feeling and had a lot of confidence in what it chose to do. It was the first time where after having extended amount of time with a model, that I felt like it was the closest to how a friend might drive you around somewhere.

If this was something like a more general training process, it definitely felt more raw and more like it wasn't constrained to be stricly in the center and it really felt like it had more of a grounded human personality.

Very cool! This may actually be my favourite exp long yet, and feels like it would benfit the most from scaling up the model and scaling out the training data

brave oriole
surreal goblet
brave oriole
odd obsidian
#

i did a blind a test today by accident, i load dtrv6 thought it was new since i saw the 12th disregarding the month lol. anyway coming from fly by wire and gwm5 it feel the same or a bit of a regress- too close to islands in street testing.

brave oriole
odd obsidian
#

im leaning towards fly by wire-its a red diff model but i dont know what it means maybe someone can explain.

trim kettle
brave oriole
last mauve
#

I find this model to have steering so sloppy and lazy that I can't tolerate it.