Hacker News Reader

Comment by AlotOfReading | Hacker News Reader

Background: I formerly worked at Cruise and specifically wrote large parts of the GNSS data pipeline.

I would be surprised if that were the cause. Fine localization should be able to resolve without GNSS input (with caveats) and there are specific tested requirements to be able to tolerate a certain number of minutes of outage even in the absence of LIDAR before position accuracy degrades beyond acceptable limits. It may be an edge case in the GNSS pipeline, but I don't think it's that edge case.

View on HN

Could this be due to bad mapping data at that particular location? Even if it was, wouldn’t they be able to pull over at a safe location with what they perceive in real time?

This account wouldn't be difficult to deanonymize, so I'll refrain from speculating. I don't know the actual cause.

I'd venture to guess the self driving stuff is a lot buggier and a lot less mature than one might think, I'd also venture to guess it's poorly tested. I remember a couple years ago watching a Cruise driver force the car to make an illegal left turn off of Mission a few years back. That was around the time that I noticed they kept struggling to deal with right turns off of Folsom as well.

My guess is that their "fail safe" is to just stop and wait for a human to move the car. I doubt they are capable of safely passing a stopped car so when one goes tits up they all get stopped. The hubris of Cruise to test these things without drivers in a busy city is pretty rich.

What was the stack used to write such a system? Was it a lot of C++? And what storage systems did they use? Sounds pretty cool.

The system as a whole used C, C++, Python, and some Java/JS-family stuff in the infrastructure backends. You can find a long tail of other languages scattered throughout the organization, like Go, Rust, Lisps, and Swift, etc. I wrote a fare-cost-calculator in Mathematica awhile ago, as an example.

As for storage systems, no idea. The stuff I know was all custom, but I was deep in the stack most of the time and don't do much with databases.

>The system as a whole used C, C++

It all makes sense now. Astounding its legal to use C/++ in a safety critical situation.

Uhh... you do realize tons of safety-critical systems in use today were written in C and C++, right? As in _most_ of them, for the past ~30 years?

Airplane systems, telecom equipment, medical equipment, satellite and space launch equipment, railway systems, etc., etc.

There's other languages used as well obviously, and even just assembly in some parts of some systems, but in terms of the majority it's C and/or C++.

Yep, I somewhat agree. The C mostly comes from Linux [1]. For the rest, "there are reasons".

[1] https://ossna2020.sched.com/event/c3QN/there-is-no-store-for...

C and C++ are used pretty extensively in the embedded world. For a lot of safety critical stuff though, typically a safer subset of those languages are used, such as MISRA C and MISRA C++, alongside static checkers and validated toolchains.

I don't know about the GNSS pipeline specifically but I used to work at Cruise, and most things at Cruise are written in C++ (although there's a smattering of Python for data/ML code that runs offboard).

For fine localization, are they using dead reckoning with steering angle and vehicle speed instead of GNSS?

Assume "fine localization" is a synonym for LIDAR throughout the AV industry. There are other sensors that are inputs to the localization algorithms, but LIDAR is typically the most important one.

I'm curious if the vehicles also use an INS system ties to an IGS.

What do you think it is if you had to guess?

Reply on news.ycombinator.com