Hacker News Reader

Comment by mcphilip | Hacker News Reader

Agreed.

My suspicion is that they are way ahead of competitors in terms of data from testing self driving cars and realize there are still so many unsolved problems that true autonomous vehicles aren't going to happen nearly as soon as competitors claim.

For instance, I see self-driving Google cars on a daily basis here in central Austin. There are a ton of zero visibility turns where you have to slowly edge out and hope no one is coming because visibility is blocked by a hedge.

How can self-driving cars handle that scenario without delegating to manual intervention?

View on HN

My question would be how can a human handle that scenario without computer assistance? You can always add a visual sensor near the front bumper of a self-driving car to safely let it peak around the corner without having to pull forward far enough to get hit.

peek

Reminds me of: https://twitter.com/stealthmountain

Oops! Too bad I can't edit :(

Mud on front bumper...hmm

How is it better to prevent the computer from safely see around the corner almost all the time only because mud will prevent it some of the time, but allow the human to continue driving while never being able to safely see around the corner?

No different from mud on a windshield. Alert goes off that it can't drive until the camera is cleaned off.

Or some sort of cleaning mechanism on the sensor. A "wiper" if you will.

So they'll automatically pull over to the side of the road, even if there's no shoulder? Or unsafe?

No, they just won't do turns with poor visibility in the meantime.

I agree with your point, but I'm not sure about your example! Can't self-driving cars "see" around the blind corner much earlier than a human could due to the fact that they have sensors on the very front of the car, as opposed to our "sensors" (eyes) back behind the windshield? They might do better!

Not only the placement, but with radars [1] you can see through a bunch of objects like trees, wooden fences, and indeed also hedges. Of course making sense of the data in a completely new environment is hard. This is where fleet learning helps a lot.

[1] This one police radar seems to also nicely work through stone http://www.camero-tech.com/product.php?ID=38

The self driving car can also react without a >=.7 second reaction time; which, at 60 mph is nearly 100 feet.

There are a lot of things that are opaque to normal-band light that are transparent to other things, like infrared, ultraviolet etc (and vice versa). This goes for more than just light. It all depends on the wavelengths.

One should not assume that simply because a human can't see through it, doesn't mean a computer with the right kinds of sensors can't.

You are not familiar with the technology. The sensors can easily "see" beyond a hedge easier than a person could. Look into it you'll see that the sensors have a far wider range of perception than a human driving.

Get a human (or some other sensor) to look around the hedge? At some point, humans are going to become "just another sensor platform" for the car's autonomy to utilize. And then we can just work to replace human sensors with cheaper sensors.

As an Austin driver I can agree with the dangers of our roads but I think those zero visibility turns are solvable in exactly the manner you propose. Slowly edge out until a front corner sensor can look down the road. I'm sure there are an unlimited other factors I can't imagine so I understand it isn't as simple as that.

Exactly. With a point cloud system, the sensor system is aware of where its sensing is blocked.

How can self-driving cars handle that scenario without delegating to manual intervention?

One possible way: http://web.mit.edu/~velten/www/corner/

I agree with your main point. Google certainly has enough data to have encountered issues beyond the progress of the automakers advancing incrementally. However, I would not be surprised if Tesla's approach is also yielding advances unavailable to Google. I would love to see their results combined at some point.

Discussing your specific example, I wonder if autonomous vehicles might learn to avoid blinds given enough crash data.

Another thought: because all this data is collected and analyzed it will be easier to automatically report this as a problem intersection to the city planners.

Granted this can happen now with people calling in but they will forget to do so, won't have images to show for backup, etc. Roads can be made safer this way.

Mercedes is working on that and it is partly available as a driving assistant called: Cross-Traffic Assist. But I assume there is still a lot of additional work to do.

Second, it will come when cars finally talk to each other, which will become mandatory in a few years.

Cars talking to each other is a terrible idea. I don't want your car to infect mine with a virus. It's bad enough that cars talk to central servers.

I agree with your sentiment. Wouldn't the scenario you posed be solved by not turning in red lights though?

A camera on a telescoping pole.

Radar?

Reply on news.ycombinator.com