Learning a Driving Simulator | Yassine Yousfi | COMMA_CON 2023
Вставка
- Опубліковано 16 вер 2024
- New low price $1250 | buy a comma 3X: comma.ai/shop/c...
openpilot can drive for hours without driver action and works on 250+ cars like Hyundai & Toyota. See if your car is compatible here: comma.ai/vehicles
How to get a job:
- comma.ai/jobs
comma links:
- comma.ai
- / comma_ai
- github.com/com...
- discord.comma.ai
Are you interested in openpilot?
Knowledge base:
- wiki.comma.ai
Check out the code:
- openpilot.comm...
Is my car supported?
- comma.ai/vehicles
Frequently Asked Questions:
- comma.ai/faq
How to setup openpilot:
- comma.ai/setup
Love that you are sharing this, despite the fact that you most likely are much further ahead on the research than what you are presenting.
Great that you are open sourcing the data as well, as a student I truly appreciate it! :)
1000$ to decrease inference latency by 50%. You can't be serious.
If it can be done in a weekend then that's $500/day.
@@britannio The info that it can be done in a weekend by somebody that isn't already familiar with the codebase is very likely a gros overestimate because if that were true one of the guys that wrote the software should be able to do it in one or two hours and there wouldn't be a reason to offer a bounty for that, because the inhouse dev would be even cheaper than 1000$. Long story short this is very likely a full month work or more.
Seriously impressive work, it's like alphago for cars.
But I'm still not convinced. Its still just a really good lanekeep. It's not smart enough to seize control and do a 180 if necessary.. I haven't seen anything about recovering from mistakes, avoiding other cars and aggressive drivers, understanding complex intersections and signs (this will be pretty much impossible), navigating road closed signs. Zipper merging. Or making split second decisions while moving. Also, there are many, many places where stop signs are invisible, or there are temporary signs which are partially visible. And there are many places where it's pretty much impossible to know where to go, or even to see the road, without Knowing and remembering the route beforehand. Most of this stuff can't be learned from videos, but can't be hard coded either. Need seriously intelligent path planning and control that has some more degrees of awareness and capability.
All these problems disappear if you just put robotaxis on a fixed route, like buses (which sort of defeats the purpose of the whole learning approach in the first place).
In terms of disengagement rate, they are still catching up to waymo, still very very far behind, and waymo has tele-assist and chase vehicles... If your goal is lets say 150,000 miles between disengagements, then You have to beat waymo many many times over, and without using tele-assist - but before they ever get to that point, they will run into problems with the path planning. I think.
Otoh, 1000 miles per disengagement (city) would already be really great compared to openpilot and tesla. It'll make a really great driving assist, especially with the ability to have human input without disengaging the system at all
Are you guys training on the reconstructed images or are you training your policy in latent space?? Id imagine pure latent space training would be much faster. Rwkv and Mamba might be also good options to decrease inference time generally but you might have tested those already
can anyone tell me what technology is comma 3 currently using
LiDAR ??
afaik 2 visible light cameras with different FOVs are on the device
@@aintgonhappen thanks dude
@@aintgonhappen and also can it see backwards or 360 coz it have only three cameras rt
Why I can’t see volvo
Because they use Flexray, same with Polestar, BMW, Mercedes
this is so behind Tesla FSD.
No, it’s Tesla that is copying Comma.
@@Wallace956 agree
agree