Hacker News Reader

Comment by sp332 | Hacker News Reader

They're not completely autonomous, but a remote driver can take control if there's an issue. If the car doesn't have a connection so a safety driver can take over, it's probably safer to brick it before it gets into a situation where it needs a safety driver.

View on HN

> They're not completely autonomous, but a remote driver can take control if there's an issue.

Waymo says they don't try to actually drive with a remote human driver, because lag. The control center can give hints, presumably along the lines of "turn around and take an alternate route", or "OK to go around obstacle", or "park and wait".

As they get more experience with these systems, they may be allowed a bit more autonomy after signal loss, enough to creep to a safe stopping position.

That is a reasonable design decision for a small fleet of test vehicles. It is not reasonable for a large fleet of production vehicles. Causing massive traffic disruptions is a safety hazard in itself, and these companies should not be allowed to scale up their fleet as large as they have until the rate of vehicle stalls is on par with normal car breakdown rates.

it seems like this incident was isolated to a single intersection in SF.

i don't think we need to make an assumption that cruise vehicles shut down when they lose cell connection. more likely is that there was a situation on this street outside a music festival that required remote intervention, and when the cars couldn't connect to get that assistance they stopped moving. and that would be a sane design decision even at scale.

I worked in the AV industry - not at Cruise but based on incidents like this and other things I’ve seen from riding in Cruise cars - they absolutely need a cell phone connection and will strand or pull over (or stop in the middle of the street) if there’s anything more than a very transient break in remote connectivity. Fundamental parts of the stack assume that if there’s any uncertainty in a situation, a remote operator will be available to clarify how the car should proceed. The car can’t make certain safety or operational guarantees without this.

Then it's not an autonomous vehicle if it needs a cell link and remote operattor. It should not stop dead in the middle of the street but follow the road code and get out of the way, like drivers with defective vehicles do.

But then that just circles back to the OP about how they're not really autonomous.

> it's probably safer to brick it before it gets into a situation where it needs a safety driver.

Bricking the device would be massive overkill. Surely they would just switch it off, or even just put it in a sleep mode.

Or… train it to handle a diverse range of freshly developing environments and scenarios without having a leash with central control.

Instead of that, just build it so it doesn't need training and can autonomously build the next generation of itself.

“Just”? Is this sarcasm? Your point fell flat. Training is required one way or another.

It's kind of an interesting idea. I'd be willing to let someone else drive me around by remote controlling my car, as long as a blaring alarm goes off whenever the network connection goes down. Some small amount of ML could be used to ease the transition, which defaults to "safely pull over and stop." That way at worst I just go back to driving normally.

Probably a terrible idea for a few reasons, but it was neat realizing there's an alternative to fully autonomous driving. Also doubt it would be legalized, but it's neat to consider.

Sporadic lag would be a killer. Think of how often you had Zoom calls that would lag for a second or two. Now, imagine the remote driver had that lag.

So design it to augment, rather than replace, a conventional autonomous driving system. If the self-driving system is kept running full-time, it can take over instantly if there's a communications dropout. That way, the car only has to stop immediately if a problematic situation happens to coincide with signal loss.

Perfection is not an option in this business, so it's time to stop demanding it. If we can manage to kill something less than 30,000 people a year on the roads, the way humans currently do, we can call it a win.

> it can take over instantly if there's a communications dropout

The only way to detect a communication dropout in a packet network is with constant exchange of keep alive packets. The detection time should be within some safety margins for the maximum allowed speed of the vehicle and the computational overheads. I think we’re speaking at least 100 pps, which is unlikely to deliver on time using public Internet over the unreliable medium.

tl;dr - I think it’s not currently possible.

Reply on news.ycombinator.com