Self Driving Cars?

[Just_Steve]

I sure hope any self driving car uses better maps than my Garmin that has brought me down several dead end streets that used to be pass thru streets about 20 years ago, and also down seasonal roads that are closed in the winter. Are self driving cars really ready for prime time?

 

[eroscott]

Are self driving cars really ready for prime time?

NO, IMHO. But driver ASSIST absolutely is. My Tesla AutoPilot version 1 (single camera vs several cameras on their version 2) is amazing to me and I don't see myself buying another long range car without it!!

There are already guys trying to defeat this warning. One said he could hang a weight on the steering wheel and it worked.

That, in a thread where an earlier guy in the same thread said having his hands on the wheel saved him when his Tesla suddenly veered towards yellow construction barrels. It has an affinity for the yellow color, hmm... See my edited post above.

It is hard to make anything fool-proof because fools are ingenious.

I think it would be rare that people are defeating it with weights or rubberbands as I saw someone test. That is Russian Roulette!

100% agree with the writing in your update. I KEEP my hands on the wheel because I can VERY quickly sense if AutoSteer/Pilot is doing something I don't expect. You can react in a fraction of a second compared to the IDIOTS with their HANDS OFF the wheel and only VISUALLY see that AutoSteer/Pilot is doing something odd. With my hands on the wheel I definitely have saved
myself from a crash. Just as in normal driving you know what situations require more defensive driving (busy road, curves, heavy traffic, cars around you not staying centered in their lane, construction, etc). I just pay additional attention to AutoSteer/Pilot in those situations.

Here is a video test showing that Tesla looks at two cars ahead. This is awesome and I've seen it used in my driving situations a few times.

Red show car 2 ahead. Yellow is 1 car ahead.

 

[NW-Bound]

I sure hope any self driving car uses better maps than my Garmin that has brought me down several dead end streets that used to be pass thru streets about 20 years ago, and also down seasonal roads that are closed in the winter. Are self driving cars really ready for prime time?

Right now, these cars only try to stay on the road. Next, they will try to avoid obstacles, bicyclists, pedestrians, dogs, cats, etc..

Only future higher-level cars will deal with the navigation to get you to where you want to be. And if they get lost, it's not something that causes loss of life and limb. I would not worry about it, only the more critical stuff like a collision or running off into a ditch.

 

[NW-Bound]

Interesting that the radar can bounce off two cars ahead. I guess the reflection has to go under the car immediately in front.

But even if it doesn't, I know that a radar-assisted braking system would enhance driving safety as it is way faster than a human driver. This is a feature that can be put on any car, and does not need to wait until an autonomous car is available.

 

[eroscott]

Right now, these cars only try to stay on the road. Next, they will try to avoid obstacles, bicyclists, pedestrians, dogs, cats, etc..

Note that Tesla is improving pedestrian recognition as well. This guy did test with different versions of the Tesla software and the latest is much improved.

ASIDE: In the past year, I've gotten about 20 OTA (over the air) updates to my Tesla while it was parked in my garage. This included new features and fixes. I've gotten a few for my Volts (recall fixes only ... no new features) ... but I've had to make dealer appointments and take my car in, drop off, pick up....

NOTE: You should watch this guys videos on a PC and use the youtube gear option to change the speed to 1.5!!!!!!!

 

[NW-Bound]

Interesting. Does Tesla announce what the update entails, or leave it up to people like these guys to do tests to reverse engineer the improvements?

 

[audreyh1]

I can't wait until we get to autonomous cars. One day many of us will no longer be safe driving and would have our licenses taken away in today's world. Self driving cars may keep us independent longer!

Sure there will still be accidents, but human error is still the biggest cause for accidents, fatal or otherwise. From what I've read, 90% of today's accidents could be avoided with self-driving cars. People who expect 0% accidents are letting the perfect be the enemy of the good - the status quo isn't better. We'd still be riding horses if your thinking prevailed. Even if self driving only reduces accidents by 50%, would you still prefer the status quo?

The tough period may be the transition between today and a completely autonomous future, when both manual and self driving cars have to co-exist.

Me too! I so look forward to the car doing most of the work and me being mostly a passenger!

 

[eroscott]

Interesting. Does Tesla announce what the update entails, or leave it up to people like these guys to do tests to reverse engineer the improvements?

Overall I would say they do a poor job of letting you know any details in the updates. You do get in-car display text telling you highlevel was was updated or added. In general, they don't clarify bug fixes but just say such and such area was improved.

The updates roll out in random order and more complicated ones start out slow until they get "feedback". First installers often post screen shots of what the update text states and post on Tesla related forums or FB groups.

ASIDE on how the OTAs work: The OTA updates provide "an update is available" notifications on your phone or the next time you are in your car. You then have to go to the car to tell it to install it immediately or at a future time and you pick the time. It is downloaded and verified for integrity before you get notificed.

 

[NW-Bound]

Thanks for the info.

About the pedestrian detection, Tesla uses only the vision camera. Google car primary sensor is the Lidar, though it also uses vision cameras and combines all the info. The Lidar detection of obstacles is nearly fool-proof, and it can see in the dark. The problem is the current spinning top is not aesthetically pleasing. Tesla claims it can do everything with vision cameras. I am very interested to see how it plays out.

PS. People drive with just their vision, so in theory vision cameras are all you need. You need them all around the car, because people turn their head when necessary. However, human acuity is better. We can read a street sign far better than what a standard camera can see. But when you combine the vision cameras with a Lidar that can see 360 degrees around the car in the dark, that's a winning combination. I do not care about sleek cars, so have no problem with a spinning top, but of course that's just me.

 

[eroscott]

Thanks for the info.

About the pedestrian detection, Tesla uses only the vision camera. Google car primary sensor is the Lidar, though it also uses vision cameras and combines all the info. The Lidar detection of obstacles is nearly fool-proof, and it can see in the dark. The problem is the current spinning top is not aesthetically pleasing. Tesla claims it can do everything with vision cameras. I am very interested to see how it plays out.

PS. People drive with just their vision, so in theory vision cameras are all you need. You need them all around the car, because people turn their head when necessary. However, human acuity is better. We can read a street sign far better than what a standard camera can see. But when you combine the vision cameras with a Lidar that can see 360 degrees around the car in the dark, that's a winning combination. I do not care about sleek cars, so have no problem with a spinning top, but of course that's just me.

The Tesla AutoPilot 2 (aka HW2 HardWare2) cars that are out have several cameras. Elon/Tesla has pointed out that the radar has a lot of capability and can see through rain, snow, darkness, etc that LIDAR is unable to. IN FACT, Bosch (radar supplier) provided Tesla with updated software that let it capture "radar image" 1000s of more times per second (as I recall) so it can detect things better and more precisely.

Radar equipment by itself, decoupled from the vehicle’s camera, can use “temporal smoothing to create a coarse point cloud, like lidar”, said Musk.

You do have a point about lidar and other car manufacturers are using it. They are apparently able to make it smaller and embed it in the bumpers or grills now. I think it is pretty expensive tho.

Image of Tesla 2nd generation hardware.

 

[NW-Bound]

I do not know if anybody has posted this video before, but this is an interesting TED talk describing Google's approach. The speaker talked about how Google decided to go for higher automation mode, because it did not believe the driver could be counted on to be alert to take over in case of trouble.
Later in the video, he showed the mapping produced by the Lidar, and how the software tracked all objects around it.

I just searched the Web and found that the speaker, Chris Urmson has left Google to form his own company. Several key engineers have left Google recently. I don't know how they can avoid using Google's trade secrets.

 

[Al in Ohio]

I would much rather we live in a country that invested in high speed rail running as efficiently as those in Western Europe ( now even China).

 

[Totoro]

Definitely!

In an earlier thread, I stated that a car driving software requires far more sophistication and faster CPU than flying an aircraft. Hence, I was both impressed at what has been done, and alarmed at the same time at the cavalier attitude I have seen. We took a lot more care, and our product was used by professional pilots, not dummies.

But isn't the margin for error much greater in a car in some respects?

You can safely stop almost any car within seconds, a tire can blowout, engine failure usually means no harm, or a broken bone and bruises. An airplane with a small defect equals death by fireball (no hyperbole there ).

Which is the reasons cars and buses have fewer (no?) redundant mechanical systems and only one driver. There is also one dimension less to handle and fewer external circumstances to take into account, like wind direction.

So it is more complex (traffic signs, other road users), but with more tolerance for mistakes?

 

[Totoro]

I don't know how they can avoid using Google's trade secrets.

They are trying lawsuits.

For brash Uber, Google's accusation of theft isn't even its worst news - LA Times

Incidentally, one of the cited reasons for splitting off and leaving apparently was quite a few employees hit big milestones and big payouts (several million I think). Suddenly they were FI!

Instead of FIRE though they set up their own shop or went elsewhere.

 

[Totoro]

No one was watching to see what kinds of drug deals or debauchery occurred in the privacy of an automatic restroom. They were quickly destroyed.

Which is why we'll have cameras in the car and app feedback. Harder to install camera surveillance in an automatic restroom.

 

[NW-Bound]

But isn't the margin for error much greater in a car in some respects?

You can safely stop almost any car within seconds, a tire can blowout, engine failure usually means no harm, or a broken bone and bruises. An airplane with a small defect equals death by fireball (no hyperbole there ).

Which is the reasons cars and buses have fewer (no?) redundant mechanical systems and only one driver. There is also one dimension less to handle and fewer external circumstances to take into account, like wind direction.

So it is more complex (traffic signs, other road users), but with more tolerance for mistakes?

Exactly! Even in airplanes, there are critical subsystems that are fault-tolerant, and less critical subsystems that are fail-passive. Fail-passive means you are allowed to detect the failure, then inhibit operation requiring that asset, and inform the pilots.

In the case of an autonomous car, if a tire blows out, the computer has to be able to detect it, then slows down and pulls off road. This is exactly what a human driver does. But the computer has to detect it. Driving on with a blown tire while the helpless human rider is screaming is simply not acceptable.

I don't know about allowing an autonomous car to miss a stop sign though. Or failure to detect a child in a crosswalk. Or a bicyclist to your right. Autonomous cars are supposed to save lives.

 

[Midpack]

In the case of an autonomous car, if a tire blows out, the computer has to be able to detect it, then slows down and pulls off road. This is exactly what a human driver does. But the computer has to detect it. Driving on with a blown tire while the helpless human rider is screaming is simply not acceptable.

I don't know about allowing an autonomous car to miss a stop sign though. Or failure to detect a child in a crosswalk. Or a bicyclist to your right. Autonomous cars are supposed to save lives.

Do you have any evidence that any of those situations are being overlooked in development of level 5 self driving cars?

 

[NW-Bound]

Do you have any evidence that any of those situations are being overlooked in development of level 5 self driving cars?

I was answering Totoro's question about some mistakes being allowed, and gave him some examples.

No, some mistakes are definitely not allowed. That's why we do not have Level 4 and 5 cars yet, right? They are still working on it, right?

 

[meierlde]

Even level 2 camera systems can detect pedestrians with the camera (My 2016 Malibu puts up the logo when it sees a pedestrian. If you added IR capability to the camera it would be easier because a human stands out a lot in IR.

 

[NW-Bound]

Even level 2 camera systems can detect pedestrians with the camera (My 2016 Malibu puts up the logo when it sees a pedestrian. If you added IR capability to the camera it would be easier because a human stands out a lot in IR.

Interesting! I wonder why Tesla's system took so long to have this capability (see earlier videos posted by eroscott), and then it does not seem reliable. Could be because they use the radar instead of the vision camera.

 

[JoeWras]

I do not follow all this, but think that true autonomous cars without V2V communications still make the road safer than it is now. The V2V provides the enhancements, such as making traffic flow more effective.

Cars travelling bumper-to-bumper and merging like zippers at 70 mph on congested freeways. Cars weaving at city street intersections without slowing down. I think people envision all this good stuff that requires V2V for coordination.

But man oh man, one hiccup and you have a chain reaction. Car tires do blow out, hardware does fail. I dunno.

I look at it a bit differently. I think V2V will be essential for the switchover period we were speaking of. Cars ahead can signal to cars back: "Manual driver jerk ahead, slow down."

And once fully autonomous: "Deer on side of road, be ready, reduce pack speed." Or: "Tire inflation warning detected on my car, slow down, be ready for evasive action." Those kind of things.

 

[Amethyst]

Now that is interesting (and alarming). Our car's lane-deviation warning (which we like, and always keep on) reacts to one particular little spot on our side street, which has no lane markings or any other kind of markings.

We have tried to figure out what could make the car beep at just that one spot. There is nothing white or shiny there. It looks just like the rest of the street, but the sensor doesn't like it for some reason known only to itself.

our MS appeared to suddenly fall in love with some yellow crash
barrels that we were going to pass. It tried to swerve toward
them, instead of just driving smoothly past in the exit-only lane.

Unknown why, because the two exit only lanes were well marked,
and we had been in the leftmost of these two long lanes for
perhaps 1000 feet, or so. Dry, clear weather, no glare, etc.

If my hands had not been on the wheel, I would have been too
late, or so it seemed.[/INDENT][/I]

 

[Totoro]

School is out, children crossing, slow down.
Roadside beggars next intersection, take next exit.
Company sponsored billboard ahead, please pay attention.


Endless possibilities

 

[NW-Bound]

I look at it a bit differently. I think V2V will be essential for the switchover period we were speaking of. Cars ahead can signal to cars back: "Manual driver jerk ahead, slow down."

And once fully autonomous: "Deer on side of road, be ready, reduce pack speed." Or: "Tire inflation warning detected on my car, slow down, be ready for evasive action." Those kind of things.

I still think these are more enhancements. Perhaps it is because I drive many miles in the country, and do not see how I can use "help" from other cars on the road. Even in the city, I avoid rush hours.

An autonomous car must be able to do at least the same as I do now, with no external aid. I don't need GPS, radar, Lidar. With just my 2 eyes, I survive so far, and same as many other drivers. I welcome enhancements, but they are not essential.

 

[NW-Bound]

Watching the video of a Tesla owner testing the car's ability to detect pedestrians should lead one to ask the obvious question: "Why is it so hard?"

I don't know if Tesla uses the vision camera or the radar to detect pedestrians.

Using the radar is tough, because of its nature. This type of radar is typically used to detect cars in front, in order to keep a safe distance, and cars are metallic. It is usually used for keeping a safe distance. While Tesla is able to make use of the scattering return from underneath the car in front to "see" another car ahead, the radar does not have a fine resolution. The nature of the radar also limits its ability to see soft targets instead of metallic objects. Hence, it is not simple to detect pedestrians with the radar.

It leads one to the next question: "Why is it hard if one has a vision camera?"

A person, even a dog or cat, shows up clearly on a vision camera particularly in broad daylight so you have to wonder why it is tough. Even Google also relies on the Lidar to augment its camera.

Programming a computer to recognise objects seen in an image is not easy. The computer must also know the distance to that object to swerve around it. To avoid killing people in a true autonomous car, it has to be extremely reliable.

Another question: "Why isn't stereo vision used to get a distance to the objects?"

Animals have evolved to have two eyes in order to judge distance to objects. The use of stereo cameras should be obvious to measure distance to objects, but nobody has found an algorithm fast enough to permit its use in real time. Again, it has to be extremely reliable for a truly autonomous operation.

Hence, while humans can drive very well using just their two eyes, car makers have to augment the vision cameras with a radar or Lidar. Both of the latter have limitations. Google's lidar has the advantage of a 360-degree view but is limited in rain and snow. Tesla has the radar that can see through rain, but only looks ahead and with a much poorer resolution, and with just close-range ultrasonic sensors to augment the side and rear cameras.

With the intrinsic limitations of these sensors, it is interesting to see how different developers build a level 4 or 5 car. They will need some very fundamental breakthroughs, as these problems are known and been worked at for a long time.