Donate to Remove ads

Got a credit card? use our Credit Card & Finance Calculators

Thanks to johnstevens77,Bhoddhisatva,scotia,Anonymous,Cornytiv34, for Donating to support the site

Musk endeavours

The Big Picture Place
odysseus2000
Lemon Half
Posts: 6363
Joined: November 8th, 2016, 11:33 pm
Has thanked: 1530 times
Been thanked: 959 times

Re: Musk endeavours

#178517

Postby odysseus2000 » November 6th, 2018, 1:26 pm

PeterGray wrote:Quote from dspp's fascinating link above.

Tesla has always been clear that Autopilot doesn’t make the car impervious to all accidents and the issues described by Thatcham won’t be a problem for drivers using Autopilot correctly.

The problem is that NO ONE will use these sorts of systems "correctly" - be they made by Tesla or anyone else.

To prevent that sort of accident you have to be hands on wheel, eyes on road, feet by pedals ALL the time, just as you would when driving "for real". No one is going to do that, particularly on a long run. (And in fact on long runs people will start falling asleep, it can be bad enough when you are driving long distances at night). There may well be a real future for these sorts of systems - but the transition is going to be dangerous - not just for people, but for manufacturers - who may well end up facing serious legal bills.


Yes, all of this is true.

But what happens if Tesla or someone else makes a self driving system that works, that is better than a human driver and leads to substantial reductions in accidents.

We as a species can say its impossible and do nothing and continue with the 3000 deaths and serious injuries on UK roads or we can try to improve things. If someone can make robotic driving work the potential pay offs in reduced human suffering and grief are gigantic as are the commercial pay offs. That is what is at stake here and I believe none of us would not want to cut road accidents and consequential human and financial suffering.

Yes, the transition may be painful and people will likely be injured and worse but the prize for anyone who can do it seems worth a lot of treasure to go after.

Regards,

BobbyD
Lemon Half
Posts: 7814
Joined: January 22nd, 2017, 2:29 pm
Has thanked: 665 times
Been thanked: 1289 times

Re: Musk endeavours

#178520

Postby BobbyD » November 6th, 2018, 1:28 pm

odysseus2000 wrote:Yes, I have read all these arguments and used to use them, but then we had Alpha-go beating world champions in many games and all of this by the algorithms self learning how to play.

Once that happened everything changed. We are no longer dealing with a dog.

Regards,


Well Google aren't... but then Waymo vehicles don't drive in to the back of stationary firetrucks.

I'm not disputing technology can't get us to full AD, I believe it will.

But tesla are still adding dogs to the pack as they make incremental upgrades to functionality.

You were talking about processor speed. Faster dogs.

tjh290633
Lemon Half
Posts: 8208
Joined: November 4th, 2016, 11:20 am
Has thanked: 913 times
Been thanked: 4096 times

Re: Musk endeavours

#178526

Postby tjh290633 » November 6th, 2018, 1:54 pm

Looking at travelling round the M25, I wonder if an autonomous car will learn to keep a safe distance from the car in front, and keep a watch of brake lights ahead in the line of traffic. It is obvious that the vast majority of drivers do not.

TJH

odysseus2000
Lemon Half
Posts: 6363
Joined: November 8th, 2016, 11:33 pm
Has thanked: 1530 times
Been thanked: 959 times

Re: Musk endeavours

#178528

Postby odysseus2000 » November 6th, 2018, 1:57 pm

BobbyD wrote:
odysseus2000 wrote:Yes, I have read all these arguments and used to use them, but then we had Alpha-go beating world champions in many games and all of this by the algorithms self learning how to play.

Once that happened everything changed. We are no longer dealing with a dog.

Regards,


Well Google aren't... but then Waymo vehicles don't drive in to the back of stationary firetrucks.

I'm not disputing technology can't get us to full AD, I believe it will.

But tesla are still adding dogs to the pack as they make incremental upgrades to functionality.

You were talking about processor speed. Faster dogs.


No, I believe the better analogy is education.

As Tesla upgrade their processors they are increasing the learning ability of their neural nets.

Whether this leads to a neural net that is super human we shall see, but even the greatest prodigies of history took a few years to be able to demonstrate their ability so we should be surprised if it takes time & learning to see if machines can drive better than people.

Regards,

odysseus2000
Lemon Half
Posts: 6363
Joined: November 8th, 2016, 11:33 pm
Has thanked: 1530 times
Been thanked: 959 times

Re: Musk endeavours

#178530

Postby odysseus2000 » November 6th, 2018, 2:00 pm

tjh290633 wrote:Looking at travelling round the M25, I wonder if an autonomous car will learn to keep a safe distance from the car in front, and keep a watch of brake lights ahead in the line of traffic. It is obvious that the vast majority of drivers do not.

TJH


Neural nets can already do all of that & look behind, sideways & front at the same time too.

The transition to full safer than human driving is incrementally just a small fraction of what has already been achieved.

Regards,

dspp
Lemon Half
Posts: 5884
Joined: November 4th, 2016, 10:53 am
Has thanked: 5825 times
Been thanked: 2127 times

Re: Musk endeavours

#178533

Postby dspp » November 6th, 2018, 2:18 pm

BobbyD wrote:
odysseus2000 wrote:A processor is likely running 64 bits at > 100 Mhz and so it has very many more cycles to process the data and react. This has always been a big selling point for AI systems, that they could operate at super human speeds.


Heard a very nice breakdown of AI the other days. If you start with a dog's intelligence you can solve some problems by making the dog think faster, some problems you can solve by using a pack of dogs and letting each handle a part of the problem, but if you want to reach human level intelligence you need to upgrade the cognitive capacity of the dog. No number of highly capable dogs will allow you to successfully emulate a human.

The human brain is the most complex known thing in the universe, it contains about 86 billion neurons each of which is connected to an average of somewhere north of 1000 other neurons, it is massively parallel and contains areas of deep specialisation whilst retaining levels of functional plasticity which allow it to come back from devastating injury, and while it is at it it creates the entire universe as you know it. The processor in a Tesla can manipulate 1's and 0's quickly, it's a pretty decent calculator. Tesla's problem isn't that their dog isn't fast enough enough, it's that it's a dog.


BD,

Nice analogy.

The processor update may be both faster dogs and more dogs, as we don't know how much parallel processing is going on. Both help.

However the software updates are more likely slowly evolving the dog into a chimp. Granted it isn't (yet) a human, but it is not just speed & parallelism. There are important changes going on in the software updates. How good they are, and what they are, I do not know (have you seen anything ?) but they are clearly making progress.

The other aspect is the sensor tech. So far Tesla are just using cameras (8x), ultrasonics (16x), and a 2D radar (1x). The interesting thing is they have not gone 3D on the radar (is the radar just doing range & bearing, or is it also doing doppler, presumably it is ? We know it is not doing elevation and as a result it could not distinguish between a bridge and a stationary car - this would appear to me to be a potentially useful sensor step), nor gone for a LIDAR. Their attitude is that they only need their chimp to see about as much as a human, not as much as an Aegis on steroids. It is of course possible that they may take a different view on this as LIDARs come down in price and form-factor, or that they will continue to take an information-light approach to the problem.

Nevertheless the reality out there is that in the hands of sensible humans who are using the machine within its stated limitations they are getting good results:
https://model3ownersclub.com/threads/au ... dark.8376/

regards, dspp

dspp
Lemon Half
Posts: 5884
Joined: November 4th, 2016, 10:53 am
Has thanked: 5825 times
Been thanked: 2127 times

Re: Musk endeavours

#178534

Postby dspp » November 6th, 2018, 2:22 pm

odysseus2000 wrote:
BobbyD wrote:
odysseus2000 wrote:Yes, I have read all these arguments and used to use them, but then we had Alpha-go beating world champions in many games and all of this by the algorithms self learning how to play.

Once that happened everything changed. We are no longer dealing with a dog.

Regards,


Well Google aren't... but then Waymo vehicles don't drive in to the back of stationary firetrucks.

I'm not disputing technology can't get us to full AD, I believe it will.

But tesla are still adding dogs to the pack as they make incremental upgrades to functionality.

You were talking about processor speed. Faster dogs.


No, I believe the better analogy is education.

As Tesla upgrade their processors they are increasing the learning ability of their neural nets.

Whether this leads to a neural net that is super human we shall see, but even the greatest prodigies of history took a few years to be able to demonstrate their ability so we should be surprised if it takes time & learning to see if machines can drive better than people.

Regards,



02000,

You are misunderstanding the process.

There is zero learning done in the car. The car is simply a data collection and processing unit. The car does no learning. The learning is done back at base in Tesla HQ. Only when each software release is ready and signed off by humans is a new set of lessons released to the cars in the field.

Imagine the lawfest if each car started learning on its own ......... (This was what I once did grad level research on, thirty years ago. This is not a new field of research.)

regards, dspp

BobbyD
Lemon Half
Posts: 7814
Joined: January 22nd, 2017, 2:29 pm
Has thanked: 665 times
Been thanked: 1289 times

Re: Musk endeavours

#178535

Postby BobbyD » November 6th, 2018, 2:23 pm

tjh290633 wrote:Looking at travelling round the M25, I wonder if an autonomous car will learn to keep a safe distance from the car in front, and keep a watch of brake lights ahead in the line of traffic. It is obvious that the vast majority of drivers do not.


Whilst I don't doubt that they will, an autonomous car that also knew within milliseconds that a car four ahead in the same lane was about to apply it's brakes and reacted accordingly would be at a significant advantage. When looking for a replacement for a human driver it's important not to restrict yourself to like for like replacement of function. Autonomous cars may not do everything as easily as human drivers, but they are capable of some incredibly powerful things which human drivers aren't. Cars can talk to other cars, they can talk to traffic signals and other infrastructure, and they can utilise what those cars and infrastructure know.

odysseus2000 wrote:
No, I believe the better analogy is education.

As Tesla upgrade their processors they are increasing the learning ability of their neural nets.


You were talking about the speed at which the processor in a Tesla can process information. That isn't education, it's a faster collision avoidance dog.

odysseus2000
Lemon Half
Posts: 6363
Joined: November 8th, 2016, 11:33 pm
Has thanked: 1530 times
Been thanked: 959 times

Re: Musk endeavours

#178539

Postby odysseus2000 » November 6th, 2018, 2:38 pm

dspp wrote:
odysseus2000 wrote:
BobbyD wrote:
Well Google aren't... but then Waymo vehicles don't drive in to the back of stationary firetrucks.

I'm not disputing technology can't get us to full AD, I believe it will.

But tesla are still adding dogs to the pack as they make incremental upgrades to functionality.

You were talking about processor speed. Faster dogs.


No, I believe the better analogy is education.

As Tesla upgrade their processors they are increasing the learning ability of their neural nets.

Whether this leads to a neural net that is super human we shall see, but even the greatest prodigies of history took a few years to be able to demonstrate their ability so we should be surprised if it takes time & learning to see if machines can drive better than people.

Regards,



02000,

You are misunderstanding the process.

There is zero learning done in the car. The car is simply a data collection and processing unit. The car does no learning. The learning is done back at base in Tesla HQ. Only when each software release is ready and signed off by humans is a new set of lessons released to the cars in the field.

Imagine the lawfest if each car started learning on its own ......... (This was what I once did grad level research on, thirty years ago. This is not a new field of research.)

regards, dspp


A car with AI has to look at the data its sensors give it and then take action based on what its neural net has been programmed to do.

The faster its processors the more data it can gather and the faster it can react.

If you want a human educational equivalent it is the child who learns geometric options and can do also sorts of things like find minima and maxima and area under curves, then the child is 'upgraded to calculus' and the child can then do the stuff done before but much more quickly.

Regards,

dspp
Lemon Half
Posts: 5884
Joined: November 4th, 2016, 10:53 am
Has thanked: 5825 times
Been thanked: 2127 times

Re: Musk endeavours

#178555

Postby dspp » November 6th, 2018, 3:37 pm

odysseus2000 wrote:
dspp wrote:
odysseus2000 wrote:

As Tesla upgrade their processors they are increasing the learning ability of their neural nets.
,



02000,

You are misunderstanding the process.

There is zero learning done in the car. The car is simply a data collection and processing unit. The car does no learning. The learning is done back at base in Tesla HQ. Only when each software release is ready and signed off by humans is a new set of lessons released to the cars in the field.

Imagine the lawfest if each car started learning on its own ......... (This was what I once did grad level research on, thirty years ago. This is not a new field of research.)

regards, dspp


A car with AI has to look at the data its sensors give it and then take action based on what its neural net has been programmed to do.

The faster its processors the more data it can gather and the faster it can react.

If you want a human educational equivalent it is the child who learns geometric options and can do also sorts of things like find minima and maxima and area under curves, then the child is 'upgraded to calculus' and the child can then do the stuff done before but much more quickly.

Regards,


No o2000, I am afraid I disagree.

Learning is different from execution. The computer in the car is just executing the program. That is what it does, and only what it does. It does no learning whatsoever in the car.

Meanwhile back at base they are downloading the data from the edge conditions and seeking to train ('learn') the neural net to better cope with them. Actually it may not only be a neural net (or multiple nets), there may be other stuff going on, but nonetheless the learning is being done back at base.

Once the learning has reached a given maturity level then a new software version is released. Since they are doing over-the-air updates all the cars with the relevant hardware enabled are getting it.

An example of this learning process is en passant being discussed here (https://forums.tesla.com/forum/forums/a ... rd-bridges). That discussion is also giving you an insight into how edge cases are explored. It it is also showing you the very real issues with false positives.

For the hardware setups see:
https://en.wikipedia.org/wiki/Tesla_Autopilot

I'm not seeing much insight into the software side on the web. But if you want to get a feel for the object recognition that is going on see the 17-min video in this link:
https://www.androidpit.com/tesla-autopilot-sensor-video

[add: see also https://teslamotorsclub.com/tmc/threads ... ux.129790/ etc ]

regards,
dspp

BobbyD
Lemon Half
Posts: 7814
Joined: January 22nd, 2017, 2:29 pm
Has thanked: 665 times
Been thanked: 1289 times

Re: Musk endeavours

#178556

Postby BobbyD » November 6th, 2018, 3:42 pm

dspp wrote:
BobbyD wrote:
odysseus2000 wrote:A processor is likely running 64 bits at > 100 Mhz and so it has very many more cycles to process the data and react. This has always been a big selling point for AI systems, that they could operate at super human speeds.


Heard a very nice breakdown of AI the other days. If you start with a dog's intelligence you can solve some problems by making the dog think faster, some problems you can solve by using a pack of dogs and letting each handle a part of the problem, but if you want to reach human level intelligence you need to upgrade the cognitive capacity of the dog. No number of highly capable dogs will allow you to successfully emulate a human.

The human brain is the most complex known thing in the universe, it contains about 86 billion neurons each of which is connected to an average of somewhere north of 1000 other neurons, it is massively parallel and contains areas of deep specialisation whilst retaining levels of functional plasticity which allow it to come back from devastating injury, and while it is at it it creates the entire universe as you know it. The processor in a Tesla can manipulate 1's and 0's quickly, it's a pretty decent calculator. Tesla's problem isn't that their dog isn't fast enough enough, it's that it's a dog.


BD,

Nice analogy.

The processor update may be both faster dogs and more dogs, as we don't know how much parallel processing is going on. Both help.

However the software updates are more likely slowly evolving the dog into a chimp. Granted it isn't (yet) a human, but it is not just speed & parallelism. There are important changes going on in the software updates. How good they are, and what they are, I do not know (have you seen anything ?) but they are clearly making progress.

The other aspect is the sensor tech. So far Tesla are just using cameras (8x), ultrasonics (16x), and a 2D radar (1x). The interesting thing is they have not gone 3D on the radar (is the radar just doing range & bearing, or is it also doing doppler, presumably it is ? We know it is not doing elevation and as a result it could not distinguish between a bridge and a stationary car - this would appear to me to be a potentially useful sensor step), nor gone for a LIDAR. Their attitude is that they only need their chimp to see about as much as a human, not as much as an Aegis on steroids. It is of course possible that they may take a different view on this as LIDARs come down in price and form-factor, or that they will continue to take an information-light approach to the problem.

Nevertheless the reality out there is that in the hands of sensible humans who are using the machine within its stated limitations they are getting good results:
https://model3ownersclub.com/threads/au ... dark.8376/

regards, dspp


Only the application is mine, the analogy itself is stolen wholesale from one podcast or another.

The great problem with this subject is the amount of information that we simply can't know about pretty much all of it. About the only thing which is reported with any degree of openness is accidents and licenses, and so little of what has been developed is available for purchase, none of which I've ridden in and five minutes on the internet is generally enough to turn up atleast one competitor I've never heard of before. Having said that I placed my first AD based bet 2 and a bit years ago and whilst the detail remains pretty blurry the shape of the landscape is somewhat clearer now then it was then. Also I keep meaning to put some decent time in to looking in to Autonomous busses...

As far as I can see Tesla are building a more widely skilled L2 machine, but I've nothing more to go on than is commonly available, less I would imagine than you given that my dislike for their approach rules Tesla out as an investment for me. It might one day beat its chimp chest and toss a bone up in to the air to the sound of also sprach zarathustra but I think some of the people who started off trying to build a chimp and have so far got quite a decent monkey on the go will already be there if they do. I'm also not entirely sure they'll ever fully eradicate some traces of dog DNA from the final creature.

Don't get me started on sensors... It seems logical to me that if you want to achieve full AD you do it with as much information as possible, and then discard that which turns out to have no value. Getting the thing to work is a world changing achievement, the bits you need to do it initially are a footnote and can be streamlined as the tech progresses. If they were planning to use LIDAR once it became more affordable wouldn't they have taken a more moderate tone? For them to take up LIDAR now would I think involve Musk backtracking on several previous statements, and what at one point appeared to be the mission statement of autopilot, and to ignore them completely with the intention of picking them up again later would seem odd.

I can certainly see the advantages in being able to differentiate between a car and a bridge... I mentioned a few posts up I think trying to achieve AD by electronically replicating the human/chimp behind the wheel is a bad way to go and I think you are right, it seems to be basically what Tesla are trying to do, relying on the bits which the human brain has been honed for over millions of years of evolution and ignoring the easy wins which computers can do without batting a silicon eyelid.

Teslas do what they do pretty well, barring the odd infatuation with stationary public service vehicles, that's not really my gripe. I don't like the grey zone between automation and human control, I don't think it's compatible with human attention patterns, and I think it's been promoted in a way which is designed to further muddy those waters. Even good users with the best of intentions aren't capable of maintaining 100% concentration during an autonomous drive, and with AD in it's infancy a few idiots can do a lot of damage. Tesla incidents have sparked 'reassessments of the future of autonomous driving' before and I still find people saying AD can't work because cars can't tell the difference between a cloud and a white van... That's not to say that what they have achieved isn't good in its own terms, I just see those terms as being somewhat less safe and more limited as a developmental jumping off point.

If they'd concentrated on really sticking electric vehicle production I'd be far more pro-Tesla. As it is I'm interested in watching the tech unfold, even from our restricted view seats and placing the odd bet on likely looking horses.

BobbyD
Lemon Half
Posts: 7814
Joined: January 22nd, 2017, 2:29 pm
Has thanked: 665 times
Been thanked: 1289 times

Re: Musk endeavours

#178563

Postby BobbyD » November 6th, 2018, 3:52 pm

dspp wrote:Learning is different from execution. The computer in the car is just executing the program. That is what it does, and only what it does. It does no learning whatsoever in the car.


A lesson learnt by the big computer back at base can be downloaded to all the little computers (cars) and be used to improve their performance. Another fantastic advantage of AD. This isn't Knight Rider.

odysseus2000
Lemon Half
Posts: 6363
Joined: November 8th, 2016, 11:33 pm
Has thanked: 1530 times
Been thanked: 959 times

Re: Musk endeavours

#178566

Postby odysseus2000 » November 6th, 2018, 4:01 pm

dspp Learning is different from execution. The computer in the car is just executing the program. That is what it does, and only what it does. It does no learning whatsoever in the car.


What I was trying to get over is that with a better processor the car has more information upon which to apply its neural net algorithms. To my way of thinking this means that the car learns more about what is happening as it drives and therefore can more correctly follow the directions of the neural net.

One could call this execution and perhaps that is a better term.

Anyhow as Arthur C Clarke put it any sufficient advanced technology is indistinguishable from magic and if Tesla can make robotic driving work it will have a near magical levitation effect on their share price.

Regards,

dspp
Lemon Half
Posts: 5884
Joined: November 4th, 2016, 10:53 am
Has thanked: 5825 times
Been thanked: 2127 times

Re: Musk endeavours

#178574

Postby dspp » November 6th, 2018, 4:36 pm

BobbyD wrote:
The great problem with this subject is the amount of information that we simply can't know about pretty much all of it. About the only thing which is reported with any degree of openness is accidents and licenses, and so little of what has been developed is available for purchase, none of which I've ridden in and five minutes on the internet is generally enough to turn up atleast one competitor I've never heard of before. Having said that I placed my first AD based bet 2 and a bit years ago and whilst the detail remains pretty blurry the shape of the landscape is somewhat clearer now then it was then. Also I keep meaning to put some decent time in to looking in to Autonomous busses...

As far as I can see Tesla are building a more widely skilled L2 machine, but I've nothing more to go on than is commonly available, less I would imagine than you given that my dislike for their approach rules Tesla out as an investment for me. It might one day beat its chimp chest and toss a bone up in to the air to the sound of also sprach zarathustra but I think some of the people who started off trying to build a chimp and have so far got quite a decent monkey on the go will already be there if they do. I'm also not entirely sure they'll ever fully eradicate some traces of dog DNA from the final creature.

Don't get me started on sensors... It seems logical to me that if you want to achieve full AD you do it with as much information as possible, and then discard that which turns out to have no value. Getting the thing to work is a world changing achievement, the bits you need to do it initially are a footnote and can be streamlined as the tech progresses. If they were planning to use LIDAR once it became more affordable wouldn't they have taken a more moderate tone? For them to take up LIDAR now would I think involve Musk backtracking on several previous statements, and what at one point appeared to be the mission statement of autopilot, and to ignore them completely with the intention of picking them up again later would seem odd.


BD,

I completely disagree with you regarding sensor fusion for AI purposes. Given the computational load in doing data extraction & then fusion it seems to me that one should choose the minimum sensor set, not the maximum set. That minimum set should have sufficient redundancy of course to be able to revert gracefully, but that is an aside. A similar aside is that the task of extraction & fusion for AI use is different than fusion for human use. Nonetheless just as a good HMI only presents the minimum necessary information so as to prevent human overload, so too should one not overdose a AI design on sensory input.

That link I put in the last post is quite a little goldmine https://teslamotorsclub.com/tmc/threads ... 790/page-4 .

That said I think autonomy was probably initially something of an afterthought for Tesla, hence their pathway. Once they realised that the market entry strategy was best done top down it then became critical. Everybody who comes from software underestimates how tightly coupled high-energy hardware systems are, and therefore their contracting strategies are generally incorrect. By combining the info we have it seems that Tesla ditched MobilEye either because of cost reasons, or because of lack of progress reasons, or to gain a core competence & associated value*. From looking at the part-deux goldmine we can see how they are analysing the environment and building it in step by step as they go to level 4/5.

If they ultimately realise that better sensors are needed then I am sure they will do it, but my guess at this stage is that vision + radar + ultrasonics will be enough.

Worth watching https://www.youtube.com/watch?time_cont ... X5xotqyRAg

regards,
dspp


*A friend of mine once declined the opportunity to become the Tesla reseller / service partner in a particular country: there are some folks one does not want to be predated on by - especially when the path from handshake to mouth is so clear. I wonder how MobilEye think about it now.

PeterGray
Lemon Slice
Posts: 847
Joined: November 4th, 2016, 11:18 am
Has thanked: 782 times
Been thanked: 343 times

Re: Musk endeavours

#178590

Postby PeterGray » November 6th, 2018, 5:18 pm

odysseus2000 wrote:But what happens if Tesla or someone else makes a self driving system that works, that is better than a human driver and leads to substantial reductions in accidents.



Sure, I don't disagree that someone, or probably several, including Tesla quite possibly, will get it right at some time. My point is that the transition is potentially dangerous, and not just to those who may be injured, but to the companies who produce the systems. In my view there is some way yet to go with both development of the systems and user education - and unfortunately, there will always be many who don't want to be educated.

PeterGray
Lemon Slice
Posts: 847
Joined: November 4th, 2016, 11:18 am
Has thanked: 782 times
Been thanked: 343 times

Re: Musk endeavours

#178592

Postby PeterGray » November 6th, 2018, 5:28 pm

dspp wrote:
PeterGray wrote:Quote from dspp's fascinating link above.

Tesla has always been clear that Autopilot doesn’t make the car impervious to all accidents and the issues described by Thatcham won’t be a problem for drivers using Autopilot correctly.

The problem is that NO ONE will use these sorts of systems "correctly" - be they made by Tesla or anyone else.

To prevent that sort of accident you have to be hands on wheel, eyes on road, feet by pedals ALL the time, just as you would when driving "for real". No one is going to do that, particularly on a long run. (And in fact on long runs people will start falling asleep, it can be bad enough when you are driving long distances at night). There may well be a real future for these sorts of systems - but the transition is going to be dangerous - not just for people, but for manufacturers - who may well end up facing serious legal bills.


Peter,

Try driving a few hundred miles using adaptive auto-braking cruise control (aka ACC), i.e. SAE-1 with human steering. You know, the sort you'll find on a new Skoda Octavia or Golf, or Nissan similar etc. I guarantee you that you would run straight into the back of the stationary vehicle in circumstances like these. This is not a Tesla-specific problem.

regards, dspp


No disagreement there. I made a very similar point about worries using plain dumb cruise control for long periods a day or so back. And nor, in anyway am I suggesting a Tesla specific problem here. But I do see semi autonomous systems raising the issues to a higher level than plain cruise control, and until someone has acceptable fully autonomous systems there are dangers not only for the occupants, but for manufacturers in the flaws that semi autonomous systems will expose. And I wouldn't describe them as "human" flaws, though Tesla and others may well try to pass them off as that. It's not a flaw, it a simple fact of human psychology that has to be accepted and dealt with, that expecting a human "driver" of a semi autonomous car to remain in full control and attention over a journey of any length is plain nonsense. We are not built like that, and putting some wording about being in charge in the manual will absolve no one when things go wrong.

odysseus2000
Lemon Half
Posts: 6363
Joined: November 8th, 2016, 11:33 pm
Has thanked: 1530 times
Been thanked: 959 times

Re: Musk endeavours

#178598

Postby odysseus2000 » November 6th, 2018, 5:46 pm

PeterGray wrote:
dspp wrote:
PeterGray wrote:Quote from dspp's fascinating link above.

Tesla has always been clear that Autopilot doesn’t make the car impervious to all accidents and the issues described by Thatcham won’t be a problem for drivers using Autopilot correctly.

The problem is that NO ONE will use these sorts of systems "correctly" - be they made by Tesla or anyone else.

To prevent that sort of accident you have to be hands on wheel, eyes on road, feet by pedals ALL the time, just as you would when driving "for real". No one is going to do that, particularly on a long run. (And in fact on long runs people will start falling asleep, it can be bad enough when you are driving long distances at night). There may well be a real future for these sorts of systems - but the transition is going to be dangerous - not just for people, but for manufacturers - who may well end up facing serious legal bills.


Peter,

Try driving a few hundred miles using adaptive auto-braking cruise control (aka ACC), i.e. SAE-1 with human steering. You know, the sort you'll find on a new Skoda Octavia or Golf, or Nissan similar etc. I guarantee you that you would run straight into the back of the stationary vehicle in circumstances like these. This is not a Tesla-specific problem.

regards, dspp


No disagreement there. I made a very similar point about worries using plain dumb cruise control for long periods a day or so back. And nor, in anyway am I suggesting a Tesla specific problem here. But I do see semi autonomous systems raising the issues to a higher level than plain cruise control, and until someone has acceptable fully autonomous systems there are dangers not only for the occupants, but for manufacturers in the flaws that semi autonomous systems will expose. And I wouldn't describe them as "human" flaws, though Tesla and others may well try to pass them off as that. It's not a flaw, it a simple fact of human psychology that has to be accepted and dealt with, that expecting a human "driver" of a semi autonomous car to remain in full control and attention over a journey of any length is plain nonsense. We are not built like that, and putting some wording about being in charge in the manual will absolve no one when things go wrong.


Yes, but its like war. If a government wants something doing they turn a blind eye to what ever the military do if it brings victory. Remember the torture of POW in the last Iraq war, yet no military folk here or in the US ended up in the slammer, ditto the police and any government organisation.

Governments recognise how import having robotic driving is and will make it so that any bad events that occur are not severely punished by the courts.

As wrong and as anti justice as that feels it is how governments operate.

Regards,

dspp
Lemon Half
Posts: 5884
Joined: November 4th, 2016, 10:53 am
Has thanked: 5825 times
Been thanked: 2127 times

Re: Musk endeavours

#178599

Postby dspp » November 6th, 2018, 5:47 pm


BobbyD
Lemon Half
Posts: 7814
Joined: January 22nd, 2017, 2:29 pm
Has thanked: 665 times
Been thanked: 1289 times

Re: Musk endeavours

#178653

Postby BobbyD » November 7th, 2018, 12:07 am

dspp wrote:I completely disagree with you regarding sensor fusion for AI purposes. Given the computational load in doing data extraction & then fusion it seems to me that one should choose the minimum sensor set, not the maximum set. That minimum set should have sufficient redundancy of course to be able to revert gracefully, but that is an aside.


I don't see the problem if you are developing a prototype which can be rationalised for production when you've actually got it working, which of course isn't what Tesla are doing. Lidar may or may not ultimately be part of the minimum requirements for AD, but the people who find out are going to be the people who first have a fully working AD system, and comparing Tesla to for example Waymo getting an unmanned trial license that doesn't look like it's going to be Tesla.

dspp wrote: By combining the info we have it seems that Tesla ditched MobilEye either because of cost reasons, or because of lack of progress reasons, or to gain a core competence & associated value*.


Tesla didn't drop MobilEye, MobilEye stopped working with Tesla citing concerns over their definition of acceptable risk following one of Tesla's fatal crashes, from memory I think it was the concrete lane separator incident.

That's not a link it's a tome...

dspp
Lemon Half
Posts: 5884
Joined: November 4th, 2016, 10:53 am
Has thanked: 5825 times
Been thanked: 2127 times

Re: Musk endeavours

#178712

Postby dspp » November 7th, 2018, 9:52 am

BD,

Re starting with the right amount of sensors I humbly disagree. Starting with far too much will simply cause dev team distraction, rework, and increase cost and time. To use an extreme example lets say a warship could have a radar, a sonar, and a opto-tracker; and the job is to create a AI-based short-range missile defense system. You'd start by ignoring the sonar data as it will simply be irrelevant for 99% of the short-range situations, and integrating it would be a very challenging job. So instead you observe that you can get a viable solution with just the two and crack on.

That tome as you call it is a real goldmine, and I spent a few hours wading through it and the various links yesterday. As best I can see the market leaders in autonomous vehicle systems are Waymo (Google/Alphabet), MobilEye (Intel), and Tesla, plus the wildcard of Apple and various Chinese. There are many interesting things about the scene, including Intel's assessment that they can sell each kit for a few thousand $ (sensors + compute + software + data services) in return for which the OEMs add $5k to the price tag; and the issues behind the MobilEye / Tesla split (which is far more complex than you suggest, imho). Oh and the tech is interesting to me given my background :)

Also interesting is the development pathway where Tesla pushed the early kit to (or beyond) its intended limits and by doing so released L2 earlier than everybody else, and in volume (450k vehicles by end Q3 2018), with the consequences we see. This in turn has caused the industry view to evolve. Previously industry saw the L1 / L2 technologies as merely being driver assist, and so whilst they developed them they were basically hobbled and put aside whilst industry pushed to get to L4/L5. But then Tesla pushed the L2 technologies to their logical limit in their HW1 using a mix of MobilEye and Tesla stuff which in practice was used both on segregated roads (aka controlled access highways with no crossing traffic) and on non-segregated roads. The spat between MobilEye and Tesla caused Tesla to switch to HW2 and rebuild their code-base, and now they are (with software 9) delivering what is generally called L2+, i.e. good enough autonomy for a lot of drivers doing a lot of driving. Interestingly MobilEye/Intel have now realised that they cannot ignore L2+ as a pathway and so they have 15-20 or so projects that are beginning to come to market to deliver a comparable functionality (e.g. GM Cadillac Super Cruise is a MobilEye system). One can niggle about which of these individual systems is best or better than the others, but they are all intended as L2 and/or L2+.

What will be interesting is how easy it is to get from L2+ to L4/5. MobilEye and Waymo think they have the true kool aid for this, and think that Tesla is in a developmental cul de sac. Reading around I think that Tesla have a good enough handle on the semantic roadscape, and on hi-res map creation, and on rules-based-driving (witness high-speed auto lane change) which are the three basic elements of the tool kit. So Tesla obviously have a competency in all three and they are getting plenty of practice in real life.

Remember of the three Tesla fatalaties everyone keeps on going on about. At least two of those were running MobilEye + Tesla on HW1. I'm not sure about the third. The first may or may not have been switched on so unclear. https://en.wikipedia.org/wiki/List_of_s ... fatalities

What is interesting is that everything coming to market in the 2019-2020 period running L2 / L2+ will be doing so without an integrated LIDAR as far as I can see. There is no industry consensus on whether a LIDAR is required for L4/L5.

It is interesting to have delved into this as it is one of the three more immediate pieces of the Tesla value proposition.

regards, dspp


Return to “Macro and Global Topics”

Who is online

Users browsing this forum: Tedx and 22 guests