Hacker News Reader

Comment by stephengillie | Hacker News Reader

Seattle can have "grey-out" conditions, where spray from other cars combines with the rainfall to dramatically limit visibility. It's rather terrifying to drive through, where you can only see for about 20 feet in any direction. Autonomous driving that entirely relies on computer vision will fail just as poorly as human eyes - multiple sensor types are necessary. How many types of sensor does this car have? I wish the article would have gone into that detail.

View on HN

Waymo recently released a large "safety report" in which it discusses the sensor suite.

https://storage.googleapis.com/sdc-prod/v1/safety-report/way...

Why would we expect computer vision to perform as poorly as human eyes? The input sensor is just part of the system - human eyes aren't very good sensors (narrow FOV, narrow set of usable wavelengths, etc) and the human brain isn't very good at detecting things (gets distracted, fooled, etc.) Anyone who's been using Google Photos search will tell you if often notices objects in the image that you didn't even register when you took the photo or viewed the image.

I'd expect the opposite - a self-driving car might outperform human drivers significantly in conditions where we tend to underperform - poor visibility, longer braking distances, poor steering control - because it will be more effective at adjusting than a human.

And maybe some kind of communication/message passing between vehicles maybe useful. For example, if a vehicle senses black ice in a certain part of the road, it can pass that information to other cars so they know to slow down when going over the same part of the road. So something like Waze but for road conditions.

> And maybe some kind of communication/message passing between vehicles maybe useful.

I've made this suggestion before, but all commenters remarked about the risks of trusting another car, trusting another signal, hackers, script kiddies, and other malicious agents. The safest inter-car communication was determined to be adaptive cruise control.

We trust traffic lights without much issue. What's the difference?

A central entity controls the traffic lights; anyone with a few grand can go buy a car.

The same central entity also heavily regulates and monitors the roads. It's a soluble problem IMO. Very difficult, but the benefits are significant.

We already trust our safety to the cars in front of us - working tail lights & signals, proper securing of rooftop loads, etc. It's just a question of whether the incremental safety of knowing about hazards outweighs the potential for intentional mischief.

Look at it this way; if you get false positives about hazards in front of you, the consequence is maybe you slow down unnecessarily. If you get a false negative, that's really no different from no data at all - your car should behave the same.

Chances are self-driving cars will have signed code and unique crypto ids stored in security enclaves. Every manufacturer will be able to see exactly which car signed the message. If someone was sending bad messages they can just ignore that id, or more likely disable the car and send the location to the police.

Sounds like a situation where the Byzantine Generals yet again rear their ugly head.

Heavy fog and rain would most likely be bad for Waymo's camera systems and its LiDAR systems, but Waymo cars also have radar, which would probably keep working pretty well. I'm not sure how well the car would perform if only radar was working well, though.

Adaptive Cruise works very well in the Seattle rain, though I haven't driven with it in a "grey-out" condition.

Are these cars using only visual data? I thought many used some kind of LIDAR as well, where that might cut through visibility issues depending on the wavelength used.

Reply on news.ycombinator.com