Hacker News Reader

Comment by bee_rider | Hacker News Reader

It couldn't possibly be GPS, right? Networks are unreliable, the cars must be able to reach a safe state if the network drops out. They appear to have stopped in the road. If these cars can't safely navigate their way to the side of the road using local sensors, they are wildly, hilariously unsuited for autonomous operation.

> at least a dozen autonomous Chevrolet Bolts from GM Cruise Automation were spotted blocking the intersection at Gough Street and Fulton Street for a couple of hours, according to reddit user seansinha.

There must be something weird about that intersection that is causing a bug.

View on HN

You've nailed the dirty little secret of the AV industry. Everyone's safety model is underpinned by the ridiculous assumption that a human will be (a) reachable with very high reliability over shitty cellular networks; (b) able to correctly interpret complex situations with extremely limited information; (c) able to resolve safety-critical situations by making planner adjustments, without direct control over steering and acceleration.

Wow. I had no idea that these things had such fundamental limitations. I'd feel a lot less safer using one now, to be honest. If you had to estimate, how long would it be before AVs would be able to do all their calculations locally? Is that something that's even under consideration at the time?

It is probably a chicken-and-egg problem. They need the data from these scenarios, so they can train the local model for later iterations accordingly.

It's a well-known thing in AV circles that some companies were actually faking their route by hardcoding them for demos etc. Also one shortcut is to use HD map data, and relying on 100k+ LIDAR sensor data that will never will be cost-effective. I know of only a handful of companies that rely on 'cheap' sensor data (cameras/radar) and use real-time calculations.

If a full time driver earns 25k-50k that's sort of equivalent to every regular taxi and uber having a $500k / $1M driving package onboard already (at 4%), right?

I don't know the industry but very naively it doesn't seem to me like $100k really breaks the unit economics? Particularly if volume / learning curve can reduce the price?

You are right in the way that the industry seems to be shifting from general-purpose self-driving to transportation/hauling.

> relying on 100k+ LIDAR sensor data that will never will be cost-effective

Eh, the $100k prices are mostly for high-end, low-production-volume research LIDARs.

There are LIDAR manufacturers claiming they can achieve prices of $500 [1] or even $200 [2] if they were fitted to every car made.

[1] https://guidehouseinsights.com/news-and-views/lower-cost-lid... [2] https://www.opsys-tech.com/post/are-we-nearing-a-shakeout-in...

Well, THAT has a Dropout-like HULU miniseries written all over it.

Tesla’s AV design philosophy differs from everyone else’s for just this reason. The idea of using centimeter-accurate models of the environment and having an always-on network connection is fundamentally at odds with handling surprises.

I don’t think a current Tesla would have a good time with no GPS either, but philosophically they stand to better handle intermittent unreliability and adverse conditions once they figure things out.

For some reason I thought these were supposed to be real products, that sounds more like they are still in the prototype/proof-of-concept phase.

> If these cars can't safely navigate their way to the side of the road using local sensors, they are wildly, hilariously unsuited for autonomous operation.

You are a lot more confident than me that this isn't the case.

Or I could be pessimistic about GPS -- these things are, albeit in limited numbers, on the road and actually being used by customers, right? I guess I'm surprised GPS is so reliable that this isn't a widespread, constant issue, if this is the normal "I lost GPS" behavior.

It could be a middle ground kind of thing. The cars have to trust something. For position you have GPS + some interpolation based on your movement sensors + some buildings/streets and maps (also unreliable since construction etc changes a lot, a lot of buildings and intersections look the same, etc).

Computers are very far away from understanding where they are from context alone. Hell most humans don't know where they are. So if you have a short GPS outage they will be able to continue navigating due to other sensors, but the longer the GPS outage/issues persist the more unlikely it is that they still know where they are.

There should definitely be a safety to drive to the side of the road and wait though.

Could it be GPS spoofing that is causing this then?

Malice, incompetence, all that jazz...

Still. I would expect some kind of sanity check so that they aren't stopping in the middle of the road.

If local sensors tell you the road is blocked, fine. But if remote data tells you one thing, and local sensors tell you something else, the fail safe should be pulling over to the side of the road.

Obviously there is no failsafe that pulls the car over, otherwise you would see that the cars had pulled over. This might be annoying on city streets but imagine if it were on the highway. An almost guaranteed 50 car pileup. If history is any guide, this will actually happen in the next few years and be followed by a lawsuit and regulation.

That's assuming that there's a space to pull over, which is often not the case on city streets. Especially, when every AV wants to pull over.

Stopping as close to a curb as possible is still better than stopping in the middle of the line. And much better than stopping on an intersection.

Also the assumption that the fail-safe _works_. Any of these fail-safes are much more complex than a classical fail-safe.

I mean, if we were driving together -- you have the steering wheel, I have the map -- and I say "woah I have no idea where we are," what would you do? I guess you would just drive around until I figured it out, or found a safe spot to pull over...

Dead reckoning works surprisingly well at least in Teslas, of which the precision can be seen when they move around in underground facilities and finally resurface: the car pops up pretty close to the entry point.

In any case, aren't all self-driving vehicles equipped with lidars and work in pre-mapped regions? I would 100% expect they are able to just use those maps with lidar to precisely geolocate themselves, probably more precisely than with GPS. And, indeed, GPS itself can at times be inaccurate or simply not available due to obstructions.

So what I'm saying, maybe it's wasn't the GPS.

> aren't all self-driving vehicles equipped with lidars

I believe Tesla has made a lot of noise about not using lidar.[0]

[0]: https://venturebeat.com/2021/07/03/tesla-ai-chief-explains-w...

In this context Tesla is not an example of a self-driving car I think, what was discussed was robo-taxis and I think all of them have lidar

Maybe at some point in the future, but Teslas aren't self-driving vehicles: you always need a driver in them, even (or maybe even: more so) in the case of FSD Beta.

I imagine these robotaxi vehicles would not have stayed on the road had they have had a safety driver in them :).

Yeah, I sort of expected a lizard brain/mammal brain split: features like object avoidance and the ability to pull over don't require the car to know where on Earth it is, and if the car can't identify to side of the road and drive to it safely with local sensors, then it just isn't fit to be on the road. That sort of stuff needs to be operate with an antilock-breaking/"fire the airbags?" level of reliability.

"Where on Earth am I" on the other hand is basically irrelevant to safety. Run it on the entertainment console or whatever. Interpolate position, come up with heuristics, whatever. I mean, it should be like two reasonably competent people on a road-trip -- if you are the driver and I'm the map guy, you aren't going to just stomp on the brakes when I inevitable lose track of where we are (I'm bad at navigating). You'd probably just go along the road while I worked it out, unless I was really confused, in which case you'd presumably pull over somewhere.

Yeah they have Lidars for 10s of 1000s of $ mounted and should be able to go for certain distances without GPS. If a GPS outage was designed as failure mode. But it would perhaps be one of the later items on any engineer‘s priority list since it is so reliable.

A network failure is, like, the world's most predictable problem though.

This is the difference between Tesla and competitors.

Tesla cars „can” drive you „safely” without connection to the network.

Some say that discard of lidar was a huge mistake by Elon, but he knows very well that this is not the way to go forward.

What does lidar have to do with connection to the network? If anything, once GPS goes down, more sensing modalities can only be helpful.

Car should be able to drive without Lidar and connection to the network.

There is currently only a single car company that accomplished that.

Why without Lidar ? Why not without camera then ? Lidar has nothing to do with Internet connection.

Reply on news.ycombinator.com