Self driving buses, trains, trucks etc

Twenty Grand wrote: »

Sure that's fair enough.

Most cars have a 20 odd year life, so decades for ubiquity is reasonable.
I also don't mind the limited human interaction. Even if you have to take control 1% of the time, that's 99% autonomous.

If the 1% of time you have to take control is not predictable, then you have to spend 100% of the journey with your hands on the wheel, paying attention to your surroundings. Not snoozing, reading, watching movies...

cnocbui wrote: »

My software writing skills aren't up to it. My son on the other hand would be well up for it, but he's working on something else at the moment. He is probably more skeptical than I am, and he writes code probably 12 hours a day, 365.

Good for him, there is a fortune to be made in IT for those with the right skills. If he is as good as you claim him to be, a job in Silicon Valley paying north of $250,000 a year awaits him.

cnocbui wrote: »

My software writing skills aren't up to it. My son on the other hand would be well up for it

...

If the target was actually self-aware general intelligence - real AI - who would be thinking that's only a decade away from full reality?

Kudos - an excellent humble-brag.:D
Am wondering about one thing. Why do you think a successful (safer than human) self driving car would need to be "self aware" to do the job? (that's kind of what that comment suggests I think?)
That seems like a very high bar.
I suppose it may involve getting closer to a sort of (limited) "general intelligence" than machines have gotten before (seeing [via sensors] + interpreting what is seen and acting on it at a human level of ability for specific task of driving a vehicle on a public road - edit: it's one job but there is quite a lot involved in it), but why self aware?

AlmightyCushion wrote: »

That's exactly what a neural net does. It learns from experience just like humans do. The neural net for self driving cars won't be based off the information and experiences of just one car, it will be based on the millions of cars they have on the road collecting data.

Alpha Go is a great example of this. It was just given the rules of the game and then played game after game. It eventually discovered effective techniques and strategies that humans had developed. Over time it even discovered ones of its own that humans now use.

A neural net can learn how to do things, only if it has encountered the circumstances before. No neural network can anticipate or predict, particularly by observing human behaviour - their sensor systems couldn't even 'see' the little girl running down the footpath, let alone understand how children behave nor anticipate or predict what could happen next.

I think general intelligence is required to match human driving capabilities. Neural nets don't come close.

Remember Musk saying he would have an autonomous car cross the US? His latest tweet:

"V9 moving to wide release now. Holding back Autopilot drive on navigation for a few more weeks of validation. Extremely difficult to achieve a general solution for self-driving that works well everywhere."

Another shining example of how just thinking something can be done with technology is faith-based rather than reality based is Musk's belief that Tesla could have factories stuffed full of robots and clever software that could make cars at blindingly fast speeds that humans couldn't begin to match - 24/7 365, even though Fiat, GM, Toyota and Nissan have all tried that. End result was Musk employing lots more people, sidelining a lot of robots and tweeting 'humans are underrated'. No kidding.

plodder wrote: »

We're only a few years from seeing self driving lorry convoys on motorways.

How many is 'a few'? Three?

How much would you bet that by the end of 2021 there will have been
- at least three convoys
- containing at least three trucks per convoy
- commercially operated (not a controlled test, but trucks carrying normal freight for an existing distributor)
- completely self-driving, no human driver required once they enter the motorway (even if the driver has hardly anything to do, or is only there 'just in case', if a driver is required in the cab ready to take over it is not a self-driving vehicle)

They would love convoy automation in Australia. As it is now, big semi's and road trains slipstream crossing the country, at least on the Nullabor bit, where one truck will drive with a bare 2m gap between it and the back of the one in front. The following truck saves a lot of fuel as it has no wind resistance. After a few hours of that, tail end charlie will overtake and take over making a hole in the air while the former lead coasts behind.

Potentially very dangerous, you would think, but they have been at it forever.

plodder wrote: »

The article below says we would have self-driving convoys by the end of this year. So, three years seems a cautious enough prediction.

https://www.theguardian.com/politics/2017/aug/25/semi-automated-truck-convoy-trials-get-uk-go-ahead-platooning

Note that the article was published in 2017 and said there would be trials in 2018, but the trials don't seem to have taken place.

RayCun wrote: »

Note that the article was published in 2017 and said there would be trials in 2018, but the trials don't seem to have taken place.

I predict they will within three years though :pac:

On the definition of "self driving" there seems to be a consensus on how it is classified with five different levels.

In a way the truck convoy is a special case, in that the convoy is led by a human driver and the following vehicles aren't truly autonomous. But, it's an interesting case, because I'm convinced we will see them before any other kind of self-driving and the idea could spread to car convoys on motorways as well, with many of the same benefits as "true" autonomous cars.

cnocbui wrote: »

Well I suppose it doesn't have to be self aware, I just suspect general intelligence might be inseparable from self awareness. Why do I think GI - because driving isn't wrote: ever observed another car, vehicle cyclist or pedestrian and read their intentions from their 'behaviour'; That 6-8 year old girl sprinting down the footpath in the direction of pedestrian crossing. No neural network is going to predict the next sequence of events. I look cars ahead often to be prepared. Anyone who thinks real driving in all weather conditions with multiple types of vehicles, objects, children pedestrians can be distilled to basically a complex but predictable scalextric set is simple minded.

A large amount of driving is entirely predictable, but the edge cases can be doozies.

We'll see I suppose.
As said many times already the technology companies think the approaches they have now or refinements/extensions of them (which presumably won't ever deliver us a self aware machine, just one very good at doing the task it has learned/been trained to do) are enough for a problem like driving + are shovelling resources at it to be "first".
Assume a lot of heads will roll + large fortunes will go down the drain if they have gotten it as badly wrong as you think...

I'm not an expert at all but as far as I understand one of the difficulties with these neural network model approaches is almost the opposite of "predictability", that they are not easily understandable for humans. It is hard to pin down exactly why they have made a particular decision (which becomes very important when they make an incorrect decision that needs to be "debugged"/fixed) and simpler models are used to try and "explain" in a rough way what the neural network is doing/why it makes the choices it does. As mentioned above (new strategies found by ones trained to play games), they have thrown up surprises for their programmers/creators when applied to problems because of this unpredictable aspect.