Hacker News Reader

Comment by jmacd | Hacker News Reader

I have a Dodge Ram. Last night I had a 400km drive to do after a very long day. I wasn't exhausted, but I certainly felt like I did not want to drive an extended period of time.

I have a Comma 3x in the truck and felt way more confident, alert and comfortable for the entire drive. OpenPilot/Sunnypilot/Frogpilot are not FSD, but they are hands off driving assistance. The 2020 Ram performs incredibly well. The latest driving models are very smooth as well, no ping-ponging and they handle passing and traffic extremely well.

A legacy car maker would be smart to acquire Comma if its for sale. They would be extremely close to a viable assisted driving capability with it.

View on HN

I used Open Pilot for ~4 years. According to connect I have 8,000 miles on Comma 2, 20,000 on Comma 3 and 2,000 on Comma 3x. I recently sold my Rav 4 and went to a Tesla. Open Pilot is actually better in a lot of ways than default Tesla auto pilot, especially because it doesn't do crazy fantom braking on freeway. Open Pilot is also way ahead of pretty much every lane assist / adaptive cruise control systems.

Obviously, FSD is way ahead of e2e open pilot with navigation, but since Open Pilot can apply very little torque to the wheel, it can't do anything gnarly. I actually trust Open pilot more at this point but I guess I just need more time with FSD. Some of that is because longitude was Toyota controlled until I used the e2e longitude model more.

Even on "chill" mode, FSD will make random quick lane changes to turn only lanes to try to get around traffic. This is 12.5.2. Even so FSD can get me from point A to B with no interventions 98% of the time.

There should be an option in FSD to have it not pass on the right and to change what speed difference it will wait to pass for (for example only pass when the car in front is [5] mph slower than what I want to go). These are separate options to look into than chill mode, and could also fall short to Comma, but thought I’d share in case you didn’t know they were there.

Mind boggling to me that a non-ping pongy lane keeping is not standard in cars. Is it standard in luxury cars? Seems like an obvious thing to add/upsell.

Non ping-ponging lane following assist is already available in many cars including KIA and Hyundai models. They're very conservative and disengage very easily. I think it's by design to minimise their legal accountability

Also Honda. In my Accord 2018 it lane kept but didn’t even play a sound when it lost tracking.

Not just legal accountability, but actual safety. They are designed so that they do not give the user a false impression of the extent of their capabilities.

Hyundai actually has two systems, LKA and LFA. LKA just tries to bounce the car back when it detects lane edges, LFA actively keeps the car in the middle of a lane.

All Hyundai models in Europe have LKA, some (more expensive) also have LFS.

I've been incredibly surprised to see that lane assist in my Kia is significantly better than that of most other (legacy non-hi-tech, think nicer hondas and lexus ICE/hybrid) cars I get a chance to drive.

I unfortunately don't have radar cruise control on my Kia, though, which would make highway driving even in traffic completely effortless, and this seems to be standard on themore expensive cars. Maybe it's for the better, though, because it does force me to be much more attentive on the road.

I am addicted to radar cruise + lane assist in my Kia. I use it all the time in traffic.

My 2019 Audi S5 was excellent at this. It would ping pong at most once then auto-correct itself to be perfectly centered in the lane.

It did some weird things like if the car in front of you was driving a bit too far to the left/right of a lane, it would copy them. Other than that it was nearly perfect, though. Never had it take an exit by accident, etc.

Their tuning on when to accelerate/brake and make it smooth needed a fair bit of work, but I found that switching the drive mode from Dynamic (Sport) to Comfort changed the eagerness of the system and smoothed things out.

> It did some weird things like if the car in front of you was driving a bit too far to the left/right of a lane, it would copy them

Wouldn't that conceptually be the right thing for the software to do, copy the human in front of it (unless it has demonstrably better information)? OT1H, "lemmings," but OTOH unless the whole line of cars were all on openpilot my life experience has been that the person in front of me by definition has more visibility than I do since their car is not blocking their view as it is mine

I am totally talking out of school, because I'm not in that space and my poor BMW chose to do its own thing[1] so it doesn't work with openpilot[2] -- although they have a dedicated #flexray channel[3] so hope springs eternal

1: https://en.wikipedia.org/wiki/FlexRay

2: https://github.com/commaai/openpilot/issues/44#issuecomment-...

3: https://discord.com/channels/469524606043160576/533838492443...

> Wouldn't that conceptually be the right thing for the software to do, copy the human in front of it (

I see people failing to follow the rules for bad reasons far more often than for any good reason. I don't want my car driving off to the side of the lane just because the car in front isn't centred. It should assume the right thing to do is to follow the rules, and hand off to me in cases that are more complicated.

Ugh. So I’m working on a fork of openpilot and the way the OP model is designed, it has its own rules that is not rooted in any legal driving framework for any state. The simple one is staying right. My state says your vehicle must stay on the right side of the road including roads without markings. OP will try to drive in the middle of the road. Another one is how OP does not distinguish people from parked cars or how oncoming cars are not tracked but simply an object the car should try to avoid (though it does not do this very well and experiences frequent disengagements due to it)

Obviously a model which manage these conditions would fair better but the comma hardware is fairly underpowered for any stronger use case.

I have added dedicated compute to my car to handle a lot of driving rules but now my solution is independent of comma. I tie into the LVDS display on the console so the integration is immersive, but it also means I don’t need comma for the hardware. My fork is also starting to diverge from OP so I may have a competing (but tangential) product!

I also notice this phenomenon in Audi. It’s as if the steering motor is applying inputs after the steering setting has been applied. So if your steering wheel is in sport mode then the motor requires additional force to turn.

I run my own forked copy of openpilot and the car cannot keep up with turns in dynamic mode. When set to comfort it can handle all turns with ease.

> A legacy car maker would be smart to acquire Comma if its for sale

My impression is that the Comma guys were never in this to sell their business

Also, unless I am misunderstanding the situation, since the code is MIT they don't need to acquire Comma to take advantage of the situation. I'd strongly suspect they all want to roll their own implementation for liability reasons, not strictly technical ones

I see that the Dodge Ram is not listed as a compatible vehicle. Could you explain how you managed to make it work?

Edit. I'm dumb. It's listed under "Ram" not "Dodge"

Reply on news.ycombinator.com