Testing Autopilot

[bwilson4web]

The investor stream for self-driving is awesome.

Bob Wilson

 

[SAronian]

Very exciting to hear the details from the hardware and software team. Excited to track the developments.

 

[Pushmi-Pullyu]

Very positive about the new computer. Sadly, that's about all I can say about the presentation that's positive. Elon is still adamant about trying to depend on cameras for almost everything.

They emphasized many times that the software still needs much development; I'm glad to see that they were being realistic about that, at least.

I don't regard Elon's timeline of true Level 4/5 self-driving within 12 months, to at most 15 months, as being even remotely realistic. There's no question he really does mean full self-driving, as he described it as being good enough that you could safely fall asleep and let the car drive itself.

I think it was about 2-3/4 years ago that Elon predicted they would have a demonstration of coast-to-coast driving by the end of that year; we're two years behind schedule on that, and counting. I consider Elon's latest prediction to be similarly over-optimistic.

The only subject on which the presentation persuaded me to change my mind, was regarding the economics of ride-sharing. In the past, I've dismissed that partly because if several people are using the car instead of just one, then the car will wear out that much faster. But Elon pointed out that objectively this would be a better use of resources; currently, automobiles typically spend about 90-95% of their time parked, doing nothing. He also claimed that the average lifespan of a car used in Tesla's ride-sharing service would be 11 years. That's much better than I expected, so if true, that would certainly make the economic case a lot more attractive.

I'm still not convinced that ride-sharing will ever be that popular. The more people use your car during a day, the greater the chance it won't be available when you need it.

I'm also utterly unconvinced that Tesla can enforce its policy of not letting anyone use a Tesla car for any ride-sharing service other than Tesla's official service, despite Elon getting rather huffy about that when challenged during a Q&A session. How would Tesla stop that? Even if they could enforce it by suing everyone who *gasp* dared to use their car with another service, how about those who buy used Tesla cars? Even if Tesla does have it in the fine print when you buy one of their cars that you agree to never join any competitor to Tesla's (hypothetical) ride-sharing service, that surely won't be enforceable against those who buy a used Tesla car. The contract with Tesla would only apply to the original buyer, not any subsequent owner.

As I see it, the worst Tesla could do would be to void the warranty for such cars, and possibly refuse to service them. And I'm not sure they could legally do even that, under some or perhaps many State laws.

​

 

[bwilson4web]

I'm reviewing the video:


I only caught snippets during the original live cast. Some of the points that 'crank my tractor' were incidental to the announced subject.

Bob Wilson

 

[bwilson4web]

I did a practice drive from Huntsville to Tunica and back using the SuperCharers in Memphis and Tupelo. Along the way, we had two more incidents:

• aborted lane change, left turn signal initiated - this is the third incident that is surprising but not really dangerous. The car begins to cross the left lane and then decides to abort and quick steers back to the original. The 'surprise' is genuine but the car handles the incident without a driving risk. Either complete the operation or try again ... looks worse than it is.
• right-side curb or lane marker turns off - Mississippi has a wide intersection that normally has a line that curves to the right with a 'dashed' line showing the true lane marker. No problem, the car handles this fine. But occasionally, the dashed line is missing and the car tries to 'split the difference' which as highway speeds, steers for the ditch. An alert driver can easily steer back into the lane. This one is more risky because car momentum to the ditch can build up fast.

Bob Wilson

 

[bwilson4web]

I forgot to include the trip costs:

• Memphis, TN Supercharge $5.56
• Memphis, TN Supercharger $1.50
• Tupelo, MS Supercharge $5.47
• Huntsville AL recharge $5.00

Total: $17.53 for 537 miles, ~= $0.0345/mile
$3.45 / 100 mile

I can live with that.

Bob Wilson

 

[Pushmi-Pullyu]

But occasionally, the dashed line is missing and the car tries to 'split the difference' which as highway speeds, steers for the ditch. An alert driver can easily steer back into the lane. This one is more risky because car momentum to the ditch can build up fast.

It's things like this which leave me very doubtful of Elon's claim that production Tesla cars will be capable of real, actual full self driving (Level 4 autonomy) within 12-15 months. I'd bet money that's not going to happen.

​

 

[bwilson4web]

As I gain experience with the basic Autopilot, especially the edge cases, I find it easier to drive the car. As for fully automated, I'm patient.

Bob Wilson

 

[bwilson4web]

We finished our first of three legs yesterday with some lesson's learned. No excitement but that 30 second reminder and steering input is a great way to increase the driver's alertness. I suspect this plays a big part in the improved safety record.

Bob Wilson

 

[bwilson4web]

Basic Autopilot

In a rush, I left my electric razor and Continuous Positive Airway Pressure (CPAP) machine at home. In a world of disposables, the razor was easily fixed (Barbasol still tastes crappy) but a CPAP is not an off the shelf item. Used to treat sleep apnea, it ensures my sleep is not interrupted by short airway blockages. Then my wife spent seven days in the hospital.

Easily confused by strangers, my presence anchors her in a safer reality to avoid more aggressive restraints. But her 24 hour care requires my sleeping in that abomination called a 'reclining chair.' It has two positions, seat parallel to floor or raised ~20 degree back. There is an extendable leg lift but the seat-to-back angle is fixed. Reclined, it is like sleeping in a bucket with the legs and knees above the rim. To mitigate, I filled the seat-to-back hole with spare sheets, blankets, and a pillow filler so I can 'cat nap' and be available to help the staff. My wife was discharged after IV antibiotics reduced the hand swelling and pain. So I took a nap at my Mom's home, packed the car, and we were off to home in Huntsville.

'Micro sleep' is a thing that my narcoleptic wife (and late aunt) suffered. Involuntarily, the victim goes to sleep for 2-5 seconds with no warning and the most effective treatment is uninterrupted sleep. After seven days cat napping in a bucket chair and one post release nap, I was primed.

Tesla's Autopilot has dynamic cruise control and automatic emergency braking to avoid running into things. The optical and ultrasonic sensor based, lane steering keeps the car from darting into a ditch or into on-coming traffic. It was on the leg to Decatur there were at least five micro sleep events that I was aware of and Autopilot kept us safe. I stopped in Decatur for a pee, stretch, and coffee break before driving the last 20 miles home.

Micro sleeps are involuntary and detected only when the individual recognizes having woken up. When driving, it becomes either a head-on collision or single car accident. Autopilot gives them a chance to survive.

Home again, I have my CPAP sleeping aid, my wife has her dogs, and me in familiar settings. Life has returned to what passes for normal ... but it is life thanks to Tesla's Autopilot.

Bob Wilson

 

[bwilson4web]

Another point of view:


Bob Wilson

 

[bwilson4web]

For thread consistency:

Source: https://insideevs.com/news/349982/nhtsa-tesla-autopilot-model-3-death-video/

Preliminary findings indicate the Autopilot system was indeed on at the time of the fatal crash involving a Model 3. The accident occurred in Delray Beach, Florida back on March 1, 2019. . . .

Missing from the article is the firmware versions and updates:

• March 1 - fatal accident occurs, Version ??
• March 26 - I took delivery of my Model 3 Standard Range Plus, Version 5.15
• April 4 - first references to Model 3, Version 8.5, in Tesla Motors Club forum

Late thought, this is the second semi-trailer crash in about 3 years. It is hard to claim this is a pattern but then the Boeing 737-MAX only had two crashes before the fleet was grounded. I am pointing out that the version of software is a problem the analysts will have to work out.

. . . The posted speed limit is 55 mph.

Preliminary data from the vehicle show that the Tesla’s Autopilot system—an advanced driver assistance system (ADAS) that provides both longitudinal and lateral control over vehicle motion—was active at the time of the crash.1 The driver engaged the Autopilot about 10 seconds before the collision. From less than 8 seconds before the crash to the time of impact, the vehicle did not detect the driver’s hands on the steering wheel. Preliminary vehicle data show that the Tesla was traveling about 68 mph when it struck the semitrailer. Neither the preliminary data nor the videos indicate that the driver or the ADAS executed evasive maneuvers.​

One clue is setting dynamic cruise control. The current version 8.5 by default sets the cruise control speed to the posted speed limit. This speed can be adjusted by operating the cruise control trim wheel apart from entering auto steer. Dynamic cruise control speed can be set before entering AutoPilot self-steering but I have seen where the car will initially limit the cruise control speed. We need to know more details about the timing of these two.

We also need to understand why the trailer did not trigger emergency stopping. My experience is the optical systems are short-range, ~30 m. The radar is the long range sensor but the following distance can vary from 1-8 (my experience.) Following trailers has been no problem but side-on was common to both fatal accidents.

Bob Wilson

 

[bwilson4web]

So this morning I ran at the side of two semi truck trailers running 8.5. One trailer without the lower aero skirt and the other with:

• At no time did the trailers show up as a shadow unless I pulled along side, parallel.
• The trailer with aero skirt would kick the car out of cruise control but no emergency braking or alarm.
• The trailer without the aero skirt was never detected except in parallel on the side.

I tried both 1 and 7 distance settings.

Bob Wilson

 

[interestedinEV]

I've been assuming that Waymo's business plan is to develop an autonomous driving system to sell to existing auto makers. Frankly I think that's a much better plan than every auto maker trying to develop one on its own. I think that's a waste of resources and money, and competing systems are unlikely to work well together. One thing that self-driving cars need to do is to share data in order to facilitate smooth traffic flow and to agree on priority for entering an intersection. If different self-driving cars use different protocols, there will almost certainly be compatibility problems.

​

You can use WYAMO to get a taxi ride in my local area. Caveat is that you cannot discuss, blog or share your experiences with others. You pay for it i.e. it is not free any more. It is a pretty busy service given the number of vans I see with passengers in the back seat. Not sure how much money they are making out of it, but people have signed up and and are using it. I know that @Pushmi-Pullyu keeps saying that Waymo should license technology and they do have a few partnerships (FCA, JLR etc.) but at this stage they are keeping everything very close to their chest. Lot of people want to go into the autonomous taxi service (Uber, Lfyt, Tesla, GM/Cruise) but the only who is actually doing it now is Waymo (albeit on a pilot scale in one market). Maybe Wyamo believes this is a viable business and wants to be in it. Maybe they do not consider that taxi in itself is not a viable business but they get so much of information on habits and preferences of people, that could not be otherwise obtained and this can be monetized. Information is money to Google. Maybe it just to prove the technology and sell it to others. Maybe it is all of them. They are holding their cards close to their chest. I would be very surprised if they do not have a fairly elaborate game plan, given that they have invested billions in this effort and appear to very deliberate and methodical.

 

[Pushmi-Pullyu]

Late thought, this is the second semi-trailer crash in about 3 years. It is hard to claim this is a pattern but then the Boeing 737-MAX only had two crashes before the fleet was grounded. I am pointing out that the version of software is a problem the analysts will have to work out.

Well said.

My first reaction to the news was "This is just what happened with the first documented fatality in a car under control of Tesla Autopilot, suggesting the tech hasn't improved in three years."

My second reaction was "I doubt this would have received any news coverage if the car was literally any car other than a Tesla."

Altho to be fair, it's Tesla's (that is, Elon Musk's) own claims which have raised expectations about the limited self-driving capabilities of the cars beyond what is reasonable or justifiable. I still think Tesla should not have used the name "Autopilot"; it suggests more functionality than is present. In fact, at one point the German government asked Tesla to change the name, but obviously Tesla hasn't complied.

Just my opinion of course, but my opinion is that this latest accident really underscores the need for a reliable SLAM system. That is, 3D scanning all around the car in real-time. Quite obviously Tesla's limited self-driving capabilities can't do that. Even during the recent "Full Self Driving" presentation, the few attempts at 3D mapping they showed were sparse and incomplete, and showed only stationary obstacles -- no moving obstacles, such as vehicles or pedestrians, at all. Still pretty far from a fully developed SLAM system, which I think (I'm fairly sure about this) is what will be required -- not just nice to have, but mandatory -- for Level 4/5 autonomy.

Active scanners are needed for real-time 3D mapping of moving obstacles. Using cameras and optical object recognition software in an attempt to "see" everything is slow (in terms of electronic speed), involves far too much computer processing, and isn't reliable enough for people to trust it with their lives.

 

[Pushmi-Pullyu]

Another point of view:


Bob Wilson

An excellent summary, and I entirely agree!

It's amazing how many people insist that we shouldn't allow self-driving cars until they've been perfected. Semi-self-driving systems are already saving lives right now! It is unfortunate, of course, that two people have lost their lives when Tesla Autopilot failed to recognize a semi trailer in the path of the car, but that doesn't erase the fact that Autopilot has doubtless already saved many, many lives, and saved untold numbers of people from serious accidents.

Demanding no fatal accidents at all is not only an impossible goal, it's an absurd one. We don't even demand that level of perfection in commercial airline flight! We do of course expect airliner manufacturers and airlines to make great efforts to reduce the number of accidents, and in fact air travel has gotten a lot safer over the past 20 years or so. But when you have people moving at high speed in relation to the ground, then sometimes fatal accidents will occur, in cars as well as airplanes. Perfect safety in high-speed travel is only possible in computer simulations; it's not possible in the real world. As the video points out, there are far too many random factors happening all the time, in the real world, for perfect safety. (Not to mention the risk of human drivers of other cars doing stupid things; something over which no self-driving car has any control!)

 

[Pushmi-Pullyu]

I know that @Pushmi-Pullyu keeps saying that Waymo should license technology and they do have a few partnerships (FCA, JLR etc.) but at this stage they are keeping everything very close to their chest.

To clarify, I meant I think that when Waymo thinks their system is sufficiently developed, then they should license it for use in existing auto makers' cars.

My impression -- and I could be mistaken, but I see a lot of evidence pointing in this direction -- my impression is that Waymo's self-driving tech is further advanced than anyone else's. However, it seems it isn't yet "ready for prime time". If it was, then passengers for Waymo's robotaxi fleet wouldn't have to sign NDAs.

I think it's a huge waste of time, resources, and money for different auto makers and other companies to be independently working toward fully functional self-driving cars. This should be treated like a Manhattan Project: A nationwide effort with many people and many companies contributing to advancement toward the goal. And once reasonably safe, reasonably well functioning autonomy is available, then all auto makers should be offered the chance to put that system into their cars.

I think the idea of having multiple competing attempts to develop self-driving automobiles is crazy. One thing that will help greatly with safety is for one autonomous car to be able to predict what another will be doing. If all autonomous cars are controlled by the same system, then not only will they be far more predictable to other autonomous cars, they will be able to share data wirelessly and seamlessly to make driving even safer, and sort out priority for lane-changing and entering intersections. That would eliminate the need for stopping at intersections; just continue to drive through, speeding up or slowing as necessary to avoid colliding with other vehicles. Similarly, vehicle-to-vehicle communications will facilitate traffic flow, helping cooperative driving rather than the stupidly competitive driving humans engage in, which should make most traffic jams a thing of the past... but only when nearly all cars are self-driving.

Vehicle-to-vehicle communications will work much better if they're all using the same system. If they are using different and incompatible systems, then the value of sharing information drops significantly, and the chance for errors in interpreting shared data increases enormously.

Human drivers are gonna continue to clog up traffic flow by foolishly fighting for the advantage of a second here and a half-second there, which inevitably results in the entire flow of traffic being slowed to very, very little faster than the slowest vehicle in the entire flow! This is one area where I think that there's a crying need for better driver education. We could teach people the advantages of driving cooperatively, rather than competitively, on multi-lane roads... but we don't.

 

[interestedinEV]

Well said.


Altho to be fair, it's Tesla's (that is, Elon Musk's) own claims which have raised expectations about the limited self-driving capabilities of the cars beyond what is reasonable or justifiable. I still think Tesla should not have used the name "Autopilot"; it suggests more functionality than is present. In fact, at one point the German government asked Tesla to change the name, but obviously Tesla hasn't complied.

While the name "Autopilot" may be misleading, the bigger issue is that Elon's penchant for the hyperbole. He has belittled his competition, told NIVDIA that they do not cut it, denigrated LIDAR. Now if he had set lower realistic expectations, it would less of a magnifying glass. He has been getting away with all of this for a very long time, one day his luck will run out.

 

[Pushmi-Pullyu]

While the name "Autopilot" may be misleading, the bigger issue is that Elon's penchant for the hyperbole. He has belittled his competition, told NIVDIA that they do not cut it, denigrated LIDAR. Now if he had set lower realistic expectations, it would less of a magnifying glass.

Sadly, I agree. Elon Musk is one of my heroes, and I think he's one of the most important innovators alive today. But his penchant for hype/ hyperbole is definitely not one of his virtues. A couple of his assertions during Tesla's "Full Self-Driving" presentation were cringe-worthy. One of those was his assertion "It is fundamentally insane to buy anything other than a Tesla."

​

 

[bwilson4web]

So I did some daylight and night test and realize forward object detection does not use radar:

• At no time does a 'ghost image' show up in front of the car, sides yes, but not in front
• Dynamic cruise control was defeated by the need for parallel lines towards the test object

In ordinary traffic, a crossing car or truck will trigger a hard brake. But why these are detected and the tractor trailer remains invisible is a mystery.

Bob Wilson