Self Driving Cars?

[Walt34]

Wow, so glad I don't live anywhere near there! Way too much traffic and too many people. The car navigation is impressive though.

 

[JoeWras]

I apologize if this Chevrolet Bolt video was already posted. I thought it was pretty impressive and there are dozens of subtle things it dealt with.

Particularly impressive was scooting over the double yellow near the end to get by that stopped delivery van. It thought about it for a while though. Nothing wrong with caution.

 

[Midpack]

But how can we be sure that the current test cars are better? Just hand waving, saying that computers are fast and technology is cool?

Just because you don't know, doesn't make it "hand waving." I know they're working on every circumstance that's been mentioned here and many, many more, I don't just assume otherwise.

 

[meierlde]

On the time constant for change. Given that the average age of a car in the US is now 11.4 years and increasing. It will likley be at least 15 years after the introduction of self driving features on moderate priced cars before any regulation on them applies on city streets. Assume a minimum of 5 years till the level 5 features appear in cars in the 30k range and you get 20+ years for any regulation.

 

[NW-Bound]

Just because you don't know, doesn't make it "hand waving." I know they're working on every circumstance that's been mentioned here and many, many more, I don't just assume otherwise.

It is true that I don't know about all the things happening, and all the different groups of people doing this.

Admittedly, I am less than impressed by the most visible project, namely Tesla, just from seeing the videos of all the mishaps on youtube.

And being an engineer, I ask myself some questions. Why does it fail in those cases? What is the cause? What will they do to fix it?

 

[NW-Bound]

I apologize if this fully autonomous driving Chevrolet Bolt video was already posted. I thought it was pretty impressive and there are dozens of subtle things it dealt with...

Yes, the software deals with the city driving very well. There's a 2nd video, and this one shows the car driving on residential streets with no lane markings. I guess if you can measure the width of the street with your sensors, you can place yourself on the proper half without lane markings.

Cruise Automation cars use Velodyne Lidars. Waymo (Google) cars have always used Lidars, and they also navigate well. No problems with detecting pedestrians there.

What is interesting with Cruise Automation cars is that they have two Lidars. Two on the roof, not one! Perhaps one is something else. I wonder how well these Lidars work in rain and snow, compared to Google Lidars.

 

[NW-Bound]

A Web site said it was indeed two Lidars, made by Velodyne. For redundancy? I like this set up already.

Here's a photo.

 

[eroscott]

Admittedly, I am less than impressed by the most visible project, namely Tesla, just from seeing the videos of all the mishaps on youtube.

I think your perspective/ruler is wrong for the AutoPilot1/HardWare1 version. As a driver assistant feature mainly intended for highway driving it is very impressive. I think there are like 1.3 billion autopilot miles now (and a few billion non-autopilot). Of my 27K miles, I probably have 17K on highways. It performs impressively there. I've driven from IL to MT last summer. A couple days were over 400 miles and would have never done that before. It is a much more relaxing (and safe) way to drive!

I've driven in some pretty good fog, snow, and rain storms and the camera and radar have works incredibly well. Beyond my expectations for sure. Nice to know it would automatically start braking even before I could. Remember unlike radar, Lidar cannot penetrate fog, heavy rain, or snow.

UPDATE: 1.3 billion https://electrek.co/2016/11/13/tesla-autopilot-billion-miles-data-self-driving-program/

As we reported before, Tesla is now gathering more data from autonomous miles driven in a day than Google’s program has logged since its inception in 2009, but the two companies are gathering fairly different datasets.

 

[Midpack]

It is true that I don't know about all the things happening, and all the different groups of people doing this.

Admittedly, I am less than impressed by the most visible project, namely Tesla, just from seeing the videos of all the mishaps on youtube.

And being an engineer, I ask myself some questions. Why does it fail in those cases? What is the cause? What will they do to fix it?

Nothing wrong with any of that.

But to keep your mistrust of Tesla in perspective, do you also ask yourself why there are 17,249 (that's 12 every minute) auto accidents with 88 fatalities per day in the US? That's the "bogey" - not perfect.

Self-driving cars have the potential to dramatically reduce accidents and fatalities, and everyone working on their development has that as a primary objective. A handful of missteps doesn't prove otherwise.

I had a long career in engineering as well. I ask myself the same questions.

https://crashstats.nhtsa.dot.gov/Api/Public/Publication/812376

 

[NW-Bound]

I think your perspective/ruler is wrong for the AutoPilot1/HardWare1 version. As a driver assistant feature mainly intended for highway driving it is very impressive. I think there are like 300 million autopilot miles now (and a few billion non-autopilot). Of my 27K miles, I probably have 17K on highways. It performs impressively there...

I dunno. The risk of a sudden twitch requiring me to react quickly to take over bothers me more than it does other people I guess.

Now, if we can have a car with both a Lidar and radar, I would feel better, having the best that money can buy. From my past background, I like to have lots of redundancy, money be damned. Aircraft cost a lot of money but the customers want safety foremost. And the risk of lawsuits and loss of business is paramount. And there's the FAA breathing down your neck.

 

[ERD50]

I'll make a prediction. By the time this stuff can be developed to handle all these various situations with high certainty, they will be obsolete. There will be other solutions, some that we can't even imagine.

Same way that people were dealing with the mess/pollution of horse manure in big cities at the beginning of the 20th century. They didn't know that in a few decades the automobile would mean we didn't need any solution to the problem - the problem went away.

That is what I think will happen. I'm not sure how/why, but I predict we won't need cars to be using detection/AI systems to try to figure out what to do. It will be done some very different way.

-ERD50

 

[ERD50]

...
Self-driving cars have the potential to dramatically reduce accidents and fatalities, ...

We don't know that. Wasn't there an earlier post with data saying that the self-driving capabilities currently require several human interventions per mile?

How much better does it need to be to be better than the average human? And how hard is it to fill that gap with technology? I really don't think those are known.

If you restate it, and say that driver assistance technology has a lot of potential to dramatically reduce accidents and fatalities, I very much agree.

-ERD50

 

[Midpack]

A little late to the thread, a lot of this has been well covered, so briefly:

Self-driving cars have the potential to dramatically reduce accidents and fatalities, and everyone working on their development has that as a primary objective. A handful of missteps doesn't prove otherwise.

We don't know that. I said potential, not know, you know the potential doesn't exist?
Wasn't there an earlier post with data saying that the self-driving capabilities currently require several human interventions per mile? There are no level 5 cars available yet, so human interventions are required, that's a given.

How much better does it need to be to be better than the average human? 90% of accidents are attributed to human error, links earlier. How much that can be reduced, we don't know yet. And how hard is it to fill that gap with technology? TBD. Early data suggested Tesla autosteer reduced accidents by 40% under some conditions, links earlier. That's just one piece of many autonomous features. I really don't think those are known. Of course not, in development, but a lot of progress so far.

If you restate it, and say that driver assistance technology has a lot of potential to dramatically reduce accidents and fatalities, I very much agree. What else would self-driving/autonomous mean?

-ERD50

 

[NW-Bound]

We don't know that. Wasn't there an earlier post with data saying that the self-driving capabilities currently require several human interventions per mile?

That was me. See post #66. It's an official report.

No, not several interventions per mile. It was one in 3 miles for Tesla, and one in 5,000 miles for Waymo.

This is the 3rd time I brought this up.

 

[ERD50]

A little late to the thread, a lot of this has been well covered, so briefly:

If you restate it, and say that driver assistance technology has a lot of potential to dramatically reduce accidents and fatalities, I very much agree.

What else would self-driving/autonomous mean?

There is a huge difference between self-driving/autonomous and driver-assistance. If you don't understand that, it might explain why you are so positive on the idea of self-driving/autonomous cars.

Self-driving/autonomous: The car takes cae of itself, w/o the driver needing to pay attention.

driver-assistance: What we have today in many cars, and it will get far better - warnings to help make the driver aware of things they might have missed. I'd especially like tech that monitors the driver to keep them involved, not lull them into complacency.

You are playing word games with 'potential' - you aren't talking just 'potential', you're talking about this tech becoming available in the foreseeable future, like our driving lifetimes.

-ERD50

 

[ERD50]

That was me. See post #66. It's an official report.

No, not several interventions per mile. It was one in 3 miles for Tesla, and one in 5,000 miles for Waymo.

This is the 3rd time I brought this up.

OK, one in 5,000 for Waymo, and one in 3 is terrible.


Even 1/5,000 - does that mean it would have had an accident w/o human intervention? And how serious? That needs to get maybe an order of magnitude better? I have not had a serious accident in over 20 years, and barely a fender bender in that time.

As mentioned, the remaining issues/conditions get progressively harder and harder to solve. Like I said, I'm skeptical that we would get there before other 'solutions' arrive, negating the need for this sort of approach.

-ERD50

 

[NW-Bound]

Nothing wrong with any of that.

But to keep your mistrust of Tesla in perspective, do you also ask yourself why there are 17,249 (that's 12 every minute) auto accidents with 88 fatalities per day in the US? That's the "bogey" - not perfect.

Self-driving cars have the potential to dramatically reduce accidents and fatalities, and everyone working on their development has that as a primary objective. A handful of missteps doesn't prove otherwise.

I had a long career in engineering as well. I ask myself the same questions.

https://crashstats.nhtsa.dot.gov/Api/Public/Publication/812376

I want to reduce accidents too, and as mentioned several times elsewhere, I do not care to drive and do not love my car like some people do. I do not even mind an ugly car with a Lidar on top.

But you bring up a good point, and make me go check some numbers. FRED site says that Americans drove a total of 3.1 trillion miles in 2015. Egads!

That works out to 8,493 million miles per day. Holy moly! But the USA is a big country. Let me double check the number by dividing it by the population of 319 million. That works out to 27 miles/day per capita. Ah, that's reasonable.

So, out of 8,494 million miles per day, Waymo cars would require human intervention 1.7 million times each day. Whoa! They need to be 100x better to match lousy humans.

PS. Now Waymo would not go out and do test drives when it rains and snows. That would bring the 5,000-mile/incidence number way down. Not good.

 

[NW-Bound]

For what it's worth, when an autopilot with autoland capability is certified with the FAA, the manufacturer has to provide analysis (FMEA - Failure Mode Effect Analysis, and FTA - Fault Tree Analysis) to show the probability of a crash is 1 in 1 billion per flight.

That's one chance of a crash in one billion flights. Now, aircraft crashes happen more frequently than that. I am not counting pilot's errors, only hardware failures. There are all kinds of things that the engineers do not foresee. But we try hard.

 

[Ronstar]

I've worked with Lidar datasets, and I'm confident that self driving cars could use such existing conditions data to drive safely in a static environment. The software should be able to handle roadway lanes, bridges, etc. IMO, the moving objects and road conditions present a great challenge to safety. These would not be in the existing conditions dataset, and the car would have to recognize and deal with these issues as they arise. Lot of variables there for software to handle.

 

[Ronstar]

A Web site said it was indeed two Lidars, made by Velodyne. For redundancy? .....[/IMG]

From what I remember, ghosting can be problematic in the data. Passing cars, pedestrians, - anything moving presents problems where the software has to decipher what is static and what is moving. The data is scrubbed to remove the ghosting. Perhaps the 2 units provide additional data to facilitate the scrubbing. And certainly redundancy wouldn't hurt.

 

[ERD50]

... But to keep your mistrust of Tesla in perspective, do you also ask yourself why there are 17,249 (that's 12 every minute) auto accidents with 88 fatalities per day in the US? That's the "bogey" - not perfect. ...

As NW-Bound just pointed out, you need the same denominator to compare. Absolute numbers aren't an adequate measure, we need to look at accidents per mile driven. Though we also can't say those earlier 'intervention' numbers would have all resulted in an accident.

And actually, that is not the bogey. You are stuck in a static model, the goal posts are moving.

Since we are getting better and better driver assistance features on cars, the relevant question is not whether a self-driving/autonomous car will have fewer accidents per mile than a current car with an average driver. The relevant question is whether a self-driving/autonomous car will have fewer accidents per mile than a car with driver assistance features, and especially one that keeps the driver involved (which I don't think we have yet - monitoring the head/eye movements as I mentioned before).

If the autonomous car is going to notify a driver when it can't figure out what to do, and that driver hasn't been paying attention, it is very likely too late for that driver to get involved look around, analyze the situation, and respond.

-ERD50

 

[explanade]

I've worked with Lidar datasets, and I'm confident that self driving cars could use such existing conditions data to drive safely in a static environment. The software should be able to handle roadway lanes, bridges, etc. IMO, the moving objects and road conditions present a great challenge to safety. These would not be in the existing conditions dataset, and the car would have to recognize and deal with these issues as they arise. Lot of variables there for software to handle.

But they're pricey.

Supposedly Waymo has found ways to reduce costs and then recently they sued Uber for patent infringement and trade secret theft.

Turns out one of their former engineers stole a lot of Waymo IP, started a new company and Uber bought that company.

 

[NW-Bound]

I've worked with Lidar datasets, and I'm confident that self driving cars could use such existing conditions data to drive safely in a static environment. The software should be able to handle roadway lanes, bridges, etc. IMO, the moving objects and road conditions present a great challenge to safety. These would not be in the existing conditions dataset, and the car would have to recognize and deal with these issues as they arise. Lot of variables there for software to handle.

Lidars are far more reliable than radar. The problem with them is that they do not work well in rain and snow.

And Waymo does not work with static conditions. They have been driving around for years, collecting data to see how to cope with moving objects, etc... More than that, they use vision cameras to read traffic signs, detect stop signs sticking out from stopped school buses.

Perhaps you have not seen the TED talk that I posted earlier. It shows how Waymo tries to address a much tougher problem than "lane following" on the freeway. I will repeat here that people have done freeway cruising long ago.

From what I remember, ghosting can be problematic in the data. Passing cars, pedestrians, - anything moving presents problems where the software has to decipher what is static and what is moving. The data is scrubbed to remove the ghosting. Perhaps the 2 units provide additional data to facilitate the scrubbing. And certainly redundancy wouldn't hurt.

Passing cars and pedestrians are not ghosts here. They are objects of great interest and importance. Waymo identifies, tracks them, and tries to predict where they are going to be.

Please watch the video whose link I posted earlier. You will understand.


PS. OK I will repeat that link here for people who are interested and want to know what is involved. Skip to about 7:30 if you are impatient and want to see recorded data of what their computer saw. This is meant for people who want to understand more about self-driving car technology, not just some demo for ordinary people.

 

[Ronstar]

Passing cars and pedestrians are not ghosts here. They are objects of interest. Waymo identifies, tracks them, and tries to predict where they are going to be..

Wow. Makes sense though for this use.

Please watch the video whose link I posted earlier. You will understand.

Thanks. - I'll watch it. Sounds interesting.

 

[ERD50]

EDIT: Ahh, I see now that the link Midpack provided has this data to 2015, I will check that and update....

https://crashstats.nhtsa.dot.gov/Api/Public/Publication/812376

checking and - methods a little different, but close enough I think. Their Table 1 and Table 2 seemed to count different, but....
---------------------------------------------------------------------------------------------
original post....

Here's some data on crashes, injuries and deaths per mile. Covers 1995/96 - 2008/09, so probably something more recent out there, but I found this and will share it while I look for more:

https://www.aaafoundation.org/sites/default/files/2012OlderDriverRisk.pdf

419.00 crashes/100 million miles, so 4.19 crashes/million miles
99.00 injuries/100 million miles, so 0.99 injuries/million miles
1.43 deaths/100 million miles, so 0.0143 deaths/million miles

Those numbers are declining (double digit %) over the time period, though I skimmed and didn't see an explanation.

-ERD50