I use sunnypilot, starpilot, and openpilot so you don't have to guess if it works. Unexpected intervention engineer 🛠
News gets road-tested, with receipts. 📋
@AC__3_3@comma_ai I actually own a 2019 M3 with FSD purchased outright. I’ve used 14 lite, recent updates have improved on the phantom acceleration and braking.
FSD v14 Lite has the best lane centering I’ve seen on this 2019 Model 3, very close to my comma four’s centering confidence.
On the first drive at 0730, the low sun washed out the cameras, so FSD wouldn’t engage.
The footage attached is my second drive, more details below 🧵
The @comma_ai Chestnut doesn't fix every turn. It does fix a lot of them currently.
Stock comma four uses a Snapdragon 845. Add the GPU and it's smoother on the ones it does get right. Still improving.
HW3 FSD updates froze. I didn’t want a monthly subscription for software that only improved on someone else’s car. Comma showed in my algorithm, people said it’s an open-source autopilot with no subscription. Despite there not being many Tesla videos with it, I bought it anyway. Still here for the same reason and why I post about it. The tinkering is just a plus.
@MRpoproppo@comma_ai Oh, Haha. I’m not sure yet, the turns are actually quite good so far. I definitely need to test out the other models a bit more before deciding.
@VagueM62565@comma_ai It works extremely well on the Model 3 and y. It's like advanced, advanced autopilot that works without lane lines. You won't regret buying it.
@BiosVova@comma_ai Comma has comma connect that serves as the main user dashboard for viewing uploaded dashcam videos, and device data. it stores the last 3 days for free. Up to a year with their optional subscription. If that's what you're asking.
They don’t seem all on the same page about whether they plan to stay at level two or are pushing towards level five. Just depends on which direction they plan on going. if they plan to push towards level five, of course a backup camera is necessary. If it's for the foreseeable future, level two, honestly, I don't see the need at the moment.
Just like a human. You never measured your own car's width either, you learned it through the windshield after enough driving. The model learned the same way, from a large stack of real driving clips showing exactly these gaps.
NEWS: CTO Harald @comma_ai put a new driving model in openpilot's latest code. Not on devices yet.
openpilot 0.11 learned by copying. This one copied too, then practiced for a score and a reward in the simulator.
Just merged a driving model into openpilot that was pre-trained on human driving, and then fine-tuned in simulation with RL.
I've wanted to train RL models that were useful in real-world robotics for a decade. So glad the time has finally come.