Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

github.com

Post reply on HN
Print view View original

21–30 of 37 posts

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#21

Post by syntaxing »

leesec wrote: post #7
syntaxing wrote: post #6

I have mix feeling about comma.ai I like what they stand for and what they're trying to achieve. But after reading their Openpilot code, it seems completely irresponsible releasing the Eon dev kit they way it is especially after their NHTSA stunt (I only reviewed the code when it first came out so things may have changed). I also wonder why this code is not being ported to something like a Jetson TX2. Especially with…

The most recent Eon isn't a OnePlus 3 it's a LeEco LePro 3. I don't see how it's really irresponsible also, it is a relatively complicated set up ( getting easier all the time ), but people definitely are aware this isn't a fully fledged product and that they should be keeping an eye at all times.

I was wondering what they were using since the Oneplus3 have been discontinued for so long.

It's irresponsible because they wrote an apk that patches to a python script... They could at least have some portion written in some hardware language like C++, especially the safety check section. Also, when I first read it there was zero redundancy. They could at a minimum do a soft watchdog how ROS does it. Putting your Drive by Wire node and the logic processing all in one unit is a horrible idea. They should have one device to do the inference like the TX2 and another do the DBW (the LePro 3).

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#22

Post by jerkstate »

iancarroll wrote: post #18
pd0wm wrote: post #9

Earlier quoted context omitted.

Can you buy them for $199 in quantity? A phone is actually a very nice and reliable piece of hardware. It has everything you need CPU, GPU, DSP, two cameras (front and back), a backup battery, a screen and wifi/lte to upload your data.

Unfortunately the code powering OpenPilot is not very reliable. As the OP alluded to it's time-sensitive code written in (very hard to understand) Python.

It's not really that time sensitive. The main event loop for something like this only has to run at like 50hz. I didn't find the code to be all that hard to understand either (at least the car interface bits needed to port to another car).

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#23

Post by rjeli »

syntaxing wrote: post #19
transistor-man wrote: post #11

Earlier quoted context omitted.

Sorry, but where is a Jetson TX2 module available for <200$?

Jetson Tx2 4GB module is expected to be [1] https://developer.nvidia.com/embedded/buy/jetson-tx2-4gb

expected, expected, expected. nvidia doesn't care about the jetson line and they keep overpromising and underdelivering.

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#25

Post by mikeash »

jerkstate wrote: post #22
iancarroll wrote: post #18

Earlier quoted context omitted.

Unfortunately the code powering OpenPilot is not very reliable. As the OP alluded to it's time-sensitive code written in (very hard to understand) Python.

It's not really that time sensitive. The main event loop for something like this only has to run at like 50hz. I didn't find the code to be all that hard to understand either (at least the car interface bits needed to port to another car).

That’s the sort of speed that has game developers, who can tolerate occasional pauses without getting anybody killed (although I’m sure some gamers act like they’re dying when it happens), avoiding memory allocations and tuning garbage collectors. That’s similar to how often the render callback for audio code will run, code where you carefully avoid taking locks or calling most system APIs and write the whole thing in C or simple C++ so implicit behaviors don’t result in audible glitches.

Python is good for many things (my whole web site runs on it) but it’s completely unsuitable for this job.

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#26

Post by bri3d »

jerkstate wrote: post #22
iancarroll wrote: post #18

Earlier quoted context omitted.

Unfortunately the code powering OpenPilot is not very reliable. As the OP alluded to it's time-sensitive code written in (very hard to understand) Python.

It's not really that time sensitive. The main event loop for something like this only has to run at like 50hz. I didn't find the code to be all that hard to understand either (at least the car interface bits needed to port to another car).

You're conflating speed and latency. The main event loop running at 50hz may or may not be fine, but a 50-100ms pause is certainly unacceptable, especially in the drive-by-wire / car interface components which are inexplicably combined with the entire platform and run on the same hardware.

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#27

Post by zucan »

ElijahLynn wrote: post #5
nharada wrote: post #4

Earlier quoted context omitted.

I really like the idea of all of this tech being open source, but I don't think it's logical that eventually it will all be 100% open. There's tons of safety critical software now that's closed source (avionics software, utility and power plant software, spacecraft control systems, etc).

I think society will eventually demand it, all it will take is a few fatalities, and even if society doesn't, I can't help but think that some company will win out _because_ they are open, like Red Hat.

Society cares about working products, not the tech behind it - I don't see calls specifically for making critical software open source any time sooner than the year of the Linux desktop.

Regulations and fines are proven tools, and even if it doesn't make autonomous cars significantly safer, it will make them expensive enough to reduce fatalities.

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#28

Post by syntaxing »

rjeli wrote: post #23
syntaxing wrote: post #19

Earlier quoted context omitted.

Jetson Tx2 4GB module is expected to be [1] https://developer.nvidia.com/embedded/buy/jetson-tx2-4gb

expected, expected, expected. nvidia doesn't care about the jetson line and they keep overpromising and underdelivering.

Hmm not too sure about using it in the industry and their over-promises. However, I did buy a TX2 to build my portfolio personally and it's been pretty fun so far. 256 CUDA cores is definitely not a lot to play with though.

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#29

Post by doctorpangloss »

beamatronic wrote: post #15
samstave wrote: post #12

Sorry for being naive on this, but I have some questions: Looking at this, it appears that they have mounted this device on a car and just had it taking in all sorts of data to effectively feed dataset into a map that can be used by self driving cars, correct? If I were doing a project like this - I would attempt to work with waste disposal companies to put these sensors on every garbage company. Garbage companies wi…

I think you’re right. Garbage collection companies, seen as a mobile sensor platform, probably can generate data (maps, wardriving, street view, AVL) worth more than their actual core business.

I think what’s more telling is the opposite implication. That despite ample sources of data, it’s not actually that valuable.

Did you ever consider that the reason that self driving cars are overhyped is because the underlying science is bad, not that there is insufficient data? Wouldn’t it be in Waymo’s interest for everyone to believe that it’s all about data, while they bide time trying to make something that works?

Re: Comma2k19 – A dataset of over 33 hours of commute in California's 280 highway

#30

Post by rjeli »

syntaxing wrote: post #28
rjeli wrote: post #23

Earlier quoted context omitted.

expected, expected, expected. nvidia doesn't care about the jetson line and they keep overpromising and underdelivering.

Hmm not too sure about using it in the industry and their over-promises. However, I did buy a TX2 to build my portfolio personally and it's been pretty fun so far. 256 CUDA cores is definitely not a lot to play with though.

the DLA makes up for the low # of cores. doesn't matter though because you can't buy them in volume
Post reply on HN