Hacker News Reader

Comment by rayiner | Hacker News Reader

I hear people say “self driving cars don’t have to be perfect, they just need to be safer than humans.” Here is a good example of why that might be harder to achieve than expected. Apparently the car had gotten confused at that exact location repeatedly before. That’s what self driving cars are going to do—if there is something “weird” it’s likely that every self driving car (at least, every one from the same manufacturer) that encounters the weird scenario will run into a problem. That could result in catastrophic failure modes at scale.

What would have happened if every car on that road had been an identical Tesla? How many crashes would have happened? How long would it have taken Tesla to issue a fix? How many miles of perfect driving would be required to make up for the cluster of crash events due to that one anomaly?

View on HN

I have a "lane sensing" Honda that errors out on the same highway exit during my commute home.

Everytime I take this same exit, my car warns me that lane sensing had an error, that collision detection has encountered an error, a few other systems have also stop working.

To play devil's advocate:

Assuming self-driving cars are connected to each other, either directly or via the internet, if one car crashes, or even just makes a minor mistake, the knowledge of how to avoid that can be transmitted to every other car on the planet.

Just as good information can be propagated, so can bad. So this had better be a robust algorithm and not one which can either be induced into something or simply learn “wrong”.

Seems unlikely that the knowledge of how to avoid making a mistake could be compiled without human analysis. The technology to do that autonomously doesn’t exist. It would almost require artificial general intelligence, and if we had that self driving would be trivial.

Automatic feature selection is the area of research you’re looking for.

The "human analysis" could be completed within 30 seconds and beamed out to all cars in the area.

No reason Tesla can't have operatives on duty 24/7 to assist and tweak/train the fleet of cars.

This is very close to what is actually happening, though not in real time.

Based on what evidence?

Based on what Tesla presented in their autonomy day event (https://www.youtube.com/watch?v=Ucp0TTmvqOE) but it's also what they've been saying publicly for a while now... in a nutshell, they ask the cars to send selected summary data back when interventions happen and in certain other cases.

Ah thank you. I had only ever heard that it was a beta feature, something they planned to implement in the future.

Tesla demonstrated a system doing exactly this at their autonomy event. The video is worth watching whatever you think Tesla.

This feels more dangerous, not less.

You get similar issues with humans though. The crossroads outside the house I lived in had restricted viability and some drivers would notice it late and crash, about one every couple of weeks for years and years. Eventually they replaced it with mini roundabouts which fixed the problem.

Also you get those epic motorway pileups in France when it gets foggy and 150+ cars crash in. Just to say we have odd failure modes too.

Reply on news.ycombinator.com