Author Topic: The autopilot, again...  (Read 1808 times)

0 Members and 1 Guest are viewing this topic.

Offline GeorgeOfTheJungleTopic starter

  • Super Contributor
  • ***
  • !
  • Posts: 2699
  • Country: tr
The autopilot, again...
« on: June 03, 2020, 01:13:34 pm »


 :clap:

Hey Elon, instead of trying to get to Mars sort your cars out here on earth first.
« Last Edit: June 03, 2020, 04:48:03 pm by GeorgeOfTheJungle »
The further a society drifts from truth, the more it will hate those who speak it.
 

Offline AlbertL

  • Regular Contributor
  • *
  • Posts: 219
  • Country: us
Re: The autopilot, again...
« Reply #1 on: June 06, 2020, 12:30:13 pm »
It seems to have a problem recognizing large light-colored obstructions.  This crash seems similar to the one where the car hit the side of a tractor-trailer (also white, I believe) that was making a left turn in front of it.   
 

Offline Siwastaja

  • Super Contributor
  • ***
  • Posts: 9327
  • Country: fi
Re: The autopilot, again...
« Reply #2 on: June 06, 2020, 12:56:42 pm »
It also failed to detect the human on the way. Which is obviously "OK", it isn't supposed to be a self-driving car. So blame the driver, as usual.

But, that horse has been beaten to death. Limiting the discussion on the technical side only, with the assumption that Tesla's target is to make a truly autonomous mode and eventually get rid of that "you are the one driving the car, don't let your hands of the wheel" disclaimer:

This is a good demonstration how miserably camera-based systems can fail. Improving cameras and image processing reduces the number of such cases, but some, even fairly common, edge cases always remain. It is technically impossible detect a color X object against color X background, while for a laser rangefinder, the color of the objects is irrelevant as long as it has enough transmittance.

You need time-of-flight of any kind to assist in such trivial cases. Even something fairly "cheap". Combining standard camera imaging with time-of-flight ranging brings the best of both worlds, but Tesla seems to really believe in using vision only. Their stance can be explained by saying that human drivers are limited in vision only, as well, so it must be possible to replicate that; but on the other hand, even if they were able to replicate the image processing human brain can do, people drive a lot of accidents due to not noticing something, so it's not a good performance target!

I thought about this when I was developing a 3D time-of-flight based mapping system which cleanly maps the existence of all floors regardless of surface materials. While doing that, I was linked Tesla PR material, supposedly showing the excellence of their 3D point clouds based on camera data. These point clouds looked great on quick glance, but did not show the streets or other even surfaces at all, only some of the objects, such as edges of buildings, vehicles, street markings... Granted, what they did show was breathtaking, with good accuracy and resolution, and a lot of points, but really the key when it comes to obstacle-avoidance (being the whole point of autonomous driving) is to understand what remains unseen. You can't afford to detect 99% of the objects, you need to detect 100%.

And for safe autonomous maneuvering, you need to both verify the surface to drive on is there - it can have a certain amount of acceptable slope -, and to verify there are no objects above that surface.

I truly believe that Tesla should allow the idea of a hybrid sensor system, adding sensing capabilities "beyond human senses", for example, distance sensing; instead of believing they can meet and even exceed the human performance by using the human senses (vision) alone. Why? Because with modern electronics, you can actually sense things human cannot, so why not take advantage of this technology? The more you can sense, the better chances you have to fuse all the sensor data together to form a complete picture; possibly with simpler algorithms. Working on a limited dataset requires really good algorithms and still misses something.

I'm sure they are working on this behind the scenes, though.
« Last Edit: June 06, 2020, 01:04:12 pm by Siwastaja »
 
The following users thanked this post: nctnico, GeorgeOfTheJungle, SiliconWizard, AlbertL

Offline AlbertL

  • Regular Contributor
  • *
  • Posts: 219
  • Country: us
Re: The autopilot, again...
« Reply #3 on: June 06, 2020, 03:26:28 pm »
I don't think fully-autonomous vehicles will be practical until three components are in place:

1) Fixed infrastructure like traffic signals capable of communicating directly with vehicles:  It seems ridiculous to expend hardware and software horsepower to visually recognize a traffic signal in a camera image, when a simple transmitter connected to the signal controller could reliably convey the signal status directly to the computers in nearby vehicles.

2) Vehicle-to-vehicle communication: in the case of the crash in the original post, the overturned truck could have immediately transmitted a warning to approaching vehicles, preventing the Tesla crash and the associated near misses by allowing the vehicles to take coordinated action to avoid the truck and each other.

3) A centralized control/coordination facility: This might be directive in nature, prescribing routes and speeds for vehicles based on their intended destinations, or just advisory, providing information but allowing the vehicle/driver to make the decisions.  Either way, it would serve as a source of real-time "big picture" information beyond what an individual vehicle can "see", allowing informed decisions that would maximize traffic throughput and safety.
     
 
The following users thanked this post: Someone, GeorgeOfTheJungle

Online nctnico

  • Super Contributor
  • ***
  • Posts: 28429
  • Country: nl
    • NCT Developments
Re: The autopilot, again...
« Reply #4 on: June 06, 2020, 03:46:43 pm »
You forgot a 4th requirement: accurate positioning so the car knows exactly where it is and doesn't have to guess based on sensor data.

Points 1 & 2 are already in service in a few places.

BTW today I went out for a bicycle ride and noticed a sign under a traffic light which says that the traffic light is connected to an app which allows cyclists to notify the traffic light they are approaching so the traffic light can plan ahead.
« Last Edit: June 06, 2020, 03:49:52 pm by nctnico »
There are small lies, big lies and then there is what is on the screen of your oscilloscope.
 
The following users thanked this post: exe

Offline SiliconWizard

  • Super Contributor
  • ***
  • Posts: 15794
  • Country: fr
Re: The autopilot, again...
« Reply #5 on: June 06, 2020, 03:47:52 pm »
I don't think fully-autonomous vehicles will be practical until three components are in place:

1) Fixed infrastructure like traffic signals capable of communicating directly with vehicles:  It seems ridiculous to expend hardware and software horsepower to visually recognize a traffic signal in a camera image, when a simple transmitter connected to the signal controller could reliably convey the signal status directly to the computers in nearby vehicles.

2) Vehicle-to-vehicle communication: in the case of the crash in the original post, the overturned truck could have immediately transmitted a warning to approaching vehicles, preventing the Tesla crash and the associated near misses by allowing the vehicles to take coordinated action to avoid the truck and each other.

3) A centralized control/coordination facility: This might be directive in nature, prescribing routes and speeds for vehicles based on their intended destinations, or just advisory, providing information but allowing the vehicle/driver to make the decisions.  Either way, it would serve as a source of real-time "big picture" information beyond what an individual vehicle can "see", allowing informed decisions that would maximize traffic throughput and safety.
   

Agree with all 3 points. That's also my view on this. Fully autonomous vehicles in normal traffic with human drivers all around and nothing to coordinate all this - it's never going to work well, and is doomed IMHO.

The current systems are "autopilots" in a similar sense as used for other kinds of vehicles, such as airplanes. They are made to automate the basics but can't be left alone unattended (knowing that of course the probability of hitting an obstacle is much lower in an airplane.)

At this point, there still needs to be a driver in CONTROL. Every such accident that I remember of was due to the driver NOT paying attention. No system is ready for this at the moment, and just because this is Tesla's (and many others) goal doesn't mean that it's ready.

Note: to the driver's "defense" - if it can be called that - I'm sure being in the driver's seat while the vehicle is in "autonomous" mode must be pretty uncomfortable, and your reflexes probably largely slowed down, so I'm not saying this is necessarily as easy as it looks to take over control of the vehicle at the right moment in case something seems to be going wrong.
« Last Edit: June 06, 2020, 03:54:03 pm by SiliconWizard »
 

Offline MaTkEOxjC

  • Contributor
  • Posts: 27
  • Country: at
Re: The autopilot, again...
« Reply #6 on: June 06, 2020, 03:58:50 pm »
I don't think fully-autonomous vehicles will be practical until three components are in place:

1) Fixed infrastructure like traffic signals capable of communicating directly with vehicles:  It seems ridiculous to expend hardware and software horsepower to visually recognize a traffic signal in a camera image, when a simple transmitter connected to the signal controller could reliably convey the signal status directly to the computers in nearby vehicles.

2) Vehicle-to-vehicle communication: in the case of the crash in the original post, the overturned truck could have immediately transmitted a warning to approaching vehicles, preventing the Tesla crash and the associated near misses by allowing the vehicles to take coordinated action to avoid the truck and each other.

3) A centralized control/coordination facility: This might be directive in nature, prescribing routes and speeds for vehicles based on their intended destinations, or just advisory, providing information but allowing the vehicle/driver to make the decisions.  Either way, it would serve as a source of real-time "big picture" information beyond what an individual vehicle can "see", allowing informed decisions that would maximize traffic throughput and safety.
   

I'm not so sure about it. Imho the car has to sense its environment as not all objects will communicate or can be expected to communicate.
There will always be a mix on the road, maybe some roads might be "autonomous only" for a limited amount of time (construction, emergency services also have to be there) but otherwise, you have to expect a not fully controllable environment.

So just think about a few factors:
  • traffic participants which do not carry electronics / communicate their position and route (cyclists, pedestrians, older cars, motorcycles, animals on the road)
  • communication may fail (someone got a jammer?)
  • infrastructure breaks sometimes, so it has so work when traffic lights are not operational e.g.
 

Offline Someone

  • Super Contributor
  • ***
  • Posts: 5154
  • Country: au
    • send complaints here
Re: The autopilot, again...
« Reply #7 on: June 06, 2020, 11:05:28 pm »
I don't think fully-autonomous vehicles will be practical until three components are in place:
There is also the larger problem of "equivalent" to a human driver, different humans have different levels of risk they will accept and a "one size fits all" autonomous driver will never keep all current drivers happy. Some will say its not safe enough, others will say its too cautious and drives too slowly.

For this case that would be slowing down for a unknown object ahead when it is uncertain if there could be a: mirage/reflection (99.9% chance), or a rolled box truck (0.001%), or other .....
 

Online nctnico

  • Super Contributor
  • ***
  • Posts: 28429
  • Country: nl
    • NCT Developments
Re: The autopilot, again...
« Reply #8 on: June 06, 2020, 11:59:49 pm »
What is the ratio of human drivers running into stationary objects versus autonomous vehicles? Yes, human drivers run their cars into stationary objects as well. Youtube is chuck full of movies showing stupid car crashes which don't involve autonomous vehicles.
« Last Edit: June 07, 2020, 12:06:19 am by nctnico »
There are small lies, big lies and then there is what is on the screen of your oscilloscope.
 
The following users thanked this post: wraper, Siwastaja

Offline Brumby

  • Supporter
  • ****
  • Posts: 12413
  • Country: au
Re: The autopilot, again...
« Reply #9 on: June 07, 2020, 06:10:20 am »
I bet the driver is glad it wasn't a concrete wall...


(Image stitched together from frames of passing video)
 

Offline vk6zgo

  • Super Contributor
  • ***
  • Posts: 7852
  • Country: au
Re: The autopilot, again...
« Reply #10 on: June 07, 2020, 06:39:19 am »
What is the ratio of human drivers running into stationary objects versus autonomous vehicles? Yes, human drivers run their cars into stationary objects as well. Youtube is chuck full of movies showing stupid car crashes which don't involve autonomous vehicles.

I think the ratio would be in the human's favour, simply because of the disparity in numbers.
There are millions of non-autonomous vehicles, compared to a mere handful of autonomous ones.

Most human drivers drive millions of kilometres over decades, without ever running into stationary objects.
The accident rate is higher for moving objects.

Another question, how many Tesla owners use the semi-autonomous system, anyway?
Most of the comments I've read were from car enthusiasts telling us how good the Teslas were to drive.
 

Offline Siwastaja

  • Super Contributor
  • ***
  • Posts: 9327
  • Country: fi
Re: The autopilot, again...
« Reply #11 on: June 07, 2020, 07:29:13 am »
I think the ratio would be in the human's favour, simply because of the disparity in numbers.
There are millions of non-autonomous vehicles, compared to a mere handful of autonomous ones.

Yes but do remember that Tesla is not an autonomous vehicle, not by any definition, and not even close. Even calling it "semi-autonomous" is hugely misleading, "semi" giving an impression of roughly around 50% of something.

It is their target, but there's a difference here. As of now, they are not even claiming they have anything close to an autonomous vehicle on market. On R&D, who knows.

As for the ratio, we don't know. Actual autonomous vehicles (all prototypes) have had very few accidents AFAIK, but the numbers on the streets are very low as well. We would need real data instead of guesses.

Until proven I'm wrong, I think that autonomous vehicle must have some kind of radar/lidar/equivalent system for 99.9999% reliable collision detection and obstacle mapping, and I think this is a hard requirement. Tesla might prove me wrong. Of course, such radar/lidar solution is the easiest part, which is exactly the reason why it should be there. I'm sure Tesla is working great on the complex AI side of things.
« Last Edit: June 07, 2020, 07:43:14 am by Siwastaja »
 

Online nctnico

  • Super Contributor
  • ***
  • Posts: 28429
  • Country: nl
    • NCT Developments
Re: The autopilot, again...
« Reply #12 on: June 07, 2020, 09:33:24 am »
What is the ratio of human drivers running into stationary objects versus autonomous vehicles? Yes, human drivers run their cars into stationary objects as well. Youtube is chuck full of movies showing stupid car crashes which don't involve autonomous vehicles.
I think the ratio would be in the human's favour, simply because of the disparity in numbers.
Without doing the actual statistics that is still wild speculation. I wouldn't be surprised that for each video on Youtube showing Tesla's auto-pilot crashing into something you can find footage of 10,000 accidents which could have easily be avoided if the driver paid attention. BTW I'm not defending Tesla here but I'm fed up with wild stories which aren't backed by solid science & facts. The world can do with much less FUD.
There are small lies, big lies and then there is what is on the screen of your oscilloscope.
 
The following users thanked this post: Siwastaja, Mr. Scram

Offline jmelson

  • Super Contributor
  • ***
  • Posts: 2851
  • Country: us
Re: The autopilot, again...
« Reply #13 on: June 07, 2020, 06:11:15 pm »
So, is the Tesla system vision only, or does it have radar, too?  I see from the crash picture that the top of the truck seems to be plastic, so maybe even a mm-wave radar would not have gotten a return off of that.

But, yes, this seems to have been a case where a vision system SHOULD have detected an obstruction.  On the other hand, 3 human-driven cars behind the Tesla did not see the truck & crashed car until VERY late and barely escaped an additional 3-4 car crash, even though they had plenty of time to see a crash ahead.

Jon
 

Offline Mr. Scram

  • Super Contributor
  • ***
  • Posts: 9821
  • Country: 00
  • Display aficionado
Re: The autopilot, again...
« Reply #14 on: June 07, 2020, 06:30:06 pm »
To be fair, the human in the Mercedes also barely noticed the wreck. I don't think it's possible to draw meaningful conclusions until you have a sound statistical understanding how Tesla's performance compares to human drivers. I think Tesla may do better than expected, based on the fact humans are utter idiots.
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf