Fully automated vehicles - 'not in our lifetime'?

GreatApe

Herzlos wrote: »

You're assuming that robo taxis will be cheaper than human drivers, not factoring in how much the robo taxi will cost

The robo taxis will be much cheaper than human taxis and eletric robo taxis will be cheaper still

A human taxi driver might buy a £20,000 taxi and use it for 5 years at 40,000 miles a year and then sell it for nearly scrap. If he earns the median male full time wage thats about £35k a year so to have a median life you want the taxi driver to have £20,000 taxi + £175,000 wages + about £20,000 fuel + say £10,000 maintenance and insurance etc = a little over 110 pence a mile

A robo eletric taxi might do 80,000 miles a year for five years before it is scrapped.
If the fleet operator charges just 50p a mile they have income of £200,000 per vehicle and £10,000 in electricity and say £10,000 in other costs. If the vehicle cost £50,000 there is a potential profit of well over £100,000 per taxi over its 5 year lifetime

And there is no reason to think a robo taxi will cost £30,000 more than a normal human car.
Tech always gets cheaper and a $200 smartphone already has more or less all the sensors needed. The huge challenge is the software but once created the marginal cost of software is virtually zero.

GreatApe

Malthusian wrote: »

After all the rubbish spoken about AI on these boards, this is the first case of someone suggesting something that could actually be feasibly be done with what is currently marketed as AI. I can imagine an AlphaZero type program being given a game to play where the gameboard is a simulation of a road network with lots of blobs moving about, and the objective is to get from any point A to B a billion times without hitting any pedestrians. DriverZero is never told which blobs are pedestrians until it hits one, in the same way as AlphaZero was never told that 1. e4 is a good idea and 1. f3 2. g4 is not. Its purpose is to run through algorithms until it finds one that allows it to correctly identify pedestrians and navigate from A to B a million times at a reasonable speed without hitting any.

The problem of course is that we have only just managed to invent an "AI" that can teach itself to play chess. Chess has 64 squares and 32 pieces which move in 6 different ways. A 3D street scene with humans milling about in all directors is an unfathomably more complex game. Whether it's feasible to have an AI which is capable of teaching itself to play "don't hit the pedestrian" in the same way that AlphaZero taught itself chess, I honestly don't know. It's quite possible that humans will figure it out before we have a powerful enough computer.

Alpha go cracked go not chess
The game go has more moves possible than the number of atoms in the universe
Versions of the alpha go have also cracked lots of different games

You also dont need to train it in a simulation, what tesla seems to be doing is just having the computer watch the human and learn from the humans driving it could then learn from the top say 10% of humans and ignore the bottom 90% which would make it better than the average human. There are a billion cars in the world doing over ten trillion miles annually. A large car company like toyota or VW could put in the software and hardware into one years production of cars and just have it watch the humans. That would give them 100 billion miles of data each year. That is 5 billion hours of human driving to learn from or equal to the computer having 500,000 years of continuous driving experience

AI learning is also highly parallel and is being expanded by magnitudes with ASIC chips
We dont need computers to be a million times faster google could just build a warehouse of a million computers and let it train on the data. The training is the complex part once trained the computation power needed to do is magnitudes less

Human cars also have a lot of downtime. A company like tesla might be able to use the 23h a day of downtime to do a massive virtual training of their system. If they had 10 million cars on the road thats 10 million networked computers that can train the AI

We are close imo

GreatApe

Arklight wrote: »

I actually agree with you that on balance the mediocre driving that AI is capable of would save more lives than the wildly inter and intra differential standards of human drivers.

In practise it cannot assess real world conditions for the same reason there is no robot on earth that can walk into your house, find your kitchen sink, find your washing up liquid and sponge, wash up a coffee cup for you, and leave.

AI sort of works with driving because roads sort of remove most of these types of variables.

At some stage we will have droids that do all that but the cars will come first because it is a $10 trillion + prize. The robot maid is trivial compared to the robot drivers

Computer vision is expanding expoentially
In specific tasks it is already super human

The Uber that killed the pedestrian in Florida had lidar yaddah yaddah. It detected her on the road, had no idea what she was because she was pushing a bike so just looked like a funny shape, and drove right into her. If Uber now reprogram their vehicles to jam on the anchors as soon as they don't recognise something on the pavement or road none of them will move an inch.

You are making a mistake non engineers make which is using the perfect as the enemy of the good. In engineering you want good not perfect because you wont get perfect. The self drive taxis can and will kill people the goal isnt to never ever kill anyone. The goal is to kill fewer people.

Also the way AI works is it gets exponentially better.
So if we start with first year the robots are just as good as humans. The second year they kill half as many. The third year half as many again and so on

There is no AI either existing now or in the pipeline that can look at a room full of people or a busy street and understand what it is seeing. You can't drive safely unless you can look at road, the condition and orientation of the other vehicles and drivers, and all the pedestrians on the sidewalk and understand what you are looking at. A decent driver on a good day can, and is legally obliged to. End of.

Sure it does not exist today but it will exist it is just information processing and it is nowhere near human level processing. A dog certainly a rat very likely or even a mouse level imagine and information processing is sufficient to do collision avoidance.

Herzlos

GreatApe wrote: »

Tech always gets cheaper and a $200 smartphone already has more or less all the sensors needed. The huge challenge is the software but once created the marginal cost of software is virtually zero.

You're orders of magnitude out - it'll be decades before a smartphone can drive a car. I'm not sure a smartphone has enough board space for all the wires, let alone the power to do anything meaningful with it. There isn't enough bandwidth to send it to the cloud either.
Like I said; the big problem currently with self driving cars is where to put and how to power all the self driving stuff.



The software will be marginal compared to the hardware costs.


Adding multiple passengers going to different destinations is turning your robo taxi into a robo bus. It's potentially more practical than a bus in that it doesn't need to go to stops if no-one needs it, or follow a set route, but that non-determinism will kill it for all but those that want the cheapest fares.


You couldn't take a robo taxi to work if you had to pick up 7 more people on the way - you've no idea when you'd arrive.


Moving freight into self driving trucks is mental too. Trains can easily pull 1000 tons of freight. You'd realistically need ~30 trucks to compensate and that's a lot of congestion and wear on roads. Electrified trains are the perfect solution for moving a large quantity of heavy stuff from one end of the country to the other.


You're also not addressing the comfort part of long distance travel. Sure you could have anything from 1-200 people in a self driving taxi, but how do you deal with the fact that they'll probably all want to pee at different points on an 8 hour journey. Have you ever tried taking a long distance bus instead of a train? It's grim.


I'll give you that cheap taxi's may kill of local train services (especially those that need to go into somewhere, change and head back), but it's not going to do anything for those that commute to a hub by train or that need to go any significant distance.



They could do wonderful things in conjunction with proper mass transport though.


And eliminating the personal car would save us so much space.

GreatApe

Herzlos wrote: »

Not really. The hardware for the test mules is in place, but it's not ready for the prime time yet, unless you're happy for most of the boot space to be computer equipment and battiries.

Another 10-20 years and you'll barely notice the processing equipment.

We have had the hardware for over a decade now

What hardware does a human use to drive? Eyes and earns

Your smartphone has a much higher resolution eye and ear than your head does
The power usage of cameras and microphones are also trivial less than a watt

You can achieve super human level driving with just cameras. Will the cars incorporate other tech like ultrasound and perhaps lidar and radar sure maybe but it is not a requirement after all you dont have ultrasound ears or radar eyes and you manage to drive a car

The processing power might be a problem ( I dont think so high end ASICs that exist today are probably already there) but even if it is computational power expands exponentially and task speciifc chips can add three magnitudes of information processing with the same types of chip manufacturing processes.

The real problem is the software and the data to feed it and crunch it doesnt exist yet. But that is mostly a matter of money and time and a $10+ trillion prize means lots of time and money is being spent on the problem

GreatApe

Herzlos wrote: »

You're orders of magnitude out - it'll be decades before a smartphone can drive a car. I'm not sure a smartphone has enough board space for all the wires, let alone the power to do anything meaningful with it. There isn't enough bandwidth to send it to the cloud either.
Like I said; the big problem currently with self driving cars is where to put and how to power all the self driving stuff.

a higher end smartphone likely already has all the hardware and processing power needed the software just does not exist. Even so by 2030 your smartphone will be at least 100x more powerful and a smartphone is a general devise an application specific computer and computer chips would be 100-1000x more powerful at a specific task than a general smartphone. So you are looking at 10,000 to 100,000 x more application specific processing power by 2030

I doubt the hardware would cost more than $1,000 the software could cost a hundred billion dollars but the hardware will be cheap (per car)

The software will be marginal compared to the hardware costs.

Sure but its all up front

Adding multiple passengers going to different destinations is turning your robo taxi into a robo bus. It's potentially more practical than a bus in that it doesn't need to go to stops if no-one needs it, or follow a set route, but that non-determinism will kill it for all but those that want the cheapest fares.

I dont think there will be many large robo buses perhaps just point to point trips like London to Birmingham. What I am getting at is its not just a typical car today. You can have everything from small cheap efficient 1 person robo taxis to larger 12 people carrier buses. Smart software plus mass usage will mean the system can pick up people starting and finishing very close to each other much better than a human bus stopping every 200 meters or a train slowing stopping waiting and accelerating two dozen times on your trip.

You couldn't take a robo taxi to work if you had to pick up 7 more people on the way - you've no idea when you'd arrive.

Why not people already take buses the london bus network carries more passangers than the tube network and you have to walk to the bus stop, wait for it and then it does 20 stops before you get to where you want to be. If that works for millions of people why wouldnt a 4 seater robo taxi that does far fewer stops not work?

Moving freight into self driving trucks is mental too. Trains can easily pull 1000 tons of freight. You'd realistically need ~30 trucks to compensate and that's a lot of congestion and wear on roads. Electrified trains are the perfect solution for moving a large quantity of heavy stuff from one end of the country to the other.

It all comes down to cost if the eletric self drive HVs are cheaper than the trains they will take over. While trains may be efficient the whole train infrastructure isnt efficient.

You're also not addressing the comfort part of long distance travel. Sure you could have anything from 1-200 people in a self driving taxi, but how do you deal with the fact that they'll probably all want to pee at different points on an 8 hour journey. Have you ever tried taking a long distance bus instead of a train? It's grim.

It wont take over all trains but in a small country like England it will take over most.
I go back to my trips I do from London to Birmingham via train it takes at least 2.5 hours from where I am to where I want to go. Why would I not take the self drive taxi which does it in 1.5 hours and would cost the same if not less assuming I book a 1 person only taxi. If I book a 4 person taxi it would cost much less than half of the train. I can also wait in my comfortable home for the taxi to pick me up rather than spend half an hour walking and busing to the train station to start my journey.

As some people leave the trains the trains will get more expensive for the rest which will have more people opt for the robo taxis as people leave the trains and their cars the fleets will become so big getting a robo taxi will take less than a minute and most journeys could be paired up with others to make things really cheap and low congestion

I'll give you that cheap taxi's may kill of local train services (especially those that need to go into somewhere, change and head back), but it's not going to do anything for those that commute to a hub by train or that need to go any significant distance.

What if you had to pay the full cost of the trains? What would the ticket costs be? Twice what they are? What if say just 20% of customers leave the trains what does that do to the remaining passengers? increase the cost of the ticket a further 25%?

Even if they were less convenient what would you go for the £150 ticket to Scotland on a train or the £30 ticket on a robo taxi that does a stop or two?

Put it this way I used to drive from the NE to London rather than use the trains. I was happy to drive all by myself than use the trains. Why would I not use a robo taxi which would be cheaper than driving myself or owning my own car. This is a 250 mile trip most uk train trips are much less than 250 miles. London to Birmingham is closer to 120 miles the taxis will take the bulk of the train trips from the midlands to the south.

They could do wonderful things in conjunction with proper mass transport though.

They will become the public transport of the future.
If you like the ideas of trains so much just pretend they are self drive trains that can go on the road and are shrunk down to 1-10 passenger size and are much lighter and overall more energy efficient and can work on the roads. :beer:

And eliminating the personal car would save us so much space.

Yes it is a huge benefit of the self drive fleets another is a much lower cost to import cars so net importers like the UK will benefit and net exporters like Japan and Germany should be scared especially if the Americans crack it

kabayiri

How do thousands of ants manage to navigate tricky routes in forests and deal with obstacles?

They dont have a supercomputer on board.

The problem becomes much easier when there is a collective approach to the problem. Our problem stems from a hybrid road network which will exist.

Herzlos

A modern smartphone just doesn't have the power to drive a car. Honestly. Look it up.


You seem to think it's a case of "if a camera can record it, we can drive" but you need to do a huge amount of processing on that visual data - object detection, speed tracking, spacial identification, for dozens if not hundreds of objects at once. You need to identify from visual queues where the road is going, what's likely to intersect you and if any deviations are required. You need to have some awareness about what's around you in all directions.



You're taking data from dozens of sensors, processing them all individually and using the complete picture to determine what to do with the car.



Your phone can probably read all the road signs in front of you but that's not enough.

westernpromise

kabayiri wrote: »

How do thousands of ants manage to navigate tricky routes in forests and deal with obstacles?

They dont have a supercomputer on board.

They don't mind if they get killed either.

kabayiri

westernpromise wrote: »

They don't mind if they get killed either.

Well, there are other places in the world which will take on the adoption risk.

Just look at how many Indian workers have died building the Qatari stadii for the world cup!

There will be places willing to try it, if the returns are there.