Autosteer on motorway = great. Autosteer on A road = dangerous?

[adsheff]

I'm thinking FSD may be a bit behind - they can't even sort out the wipers!

Seriously though, even if they make it intelligent enough, having a car drive itself would seem to me impossible without multiple redundant cameras / sensors. How many times have we had the "Blindspot detection limited", "Lane Departure Will Not Function" messages because the cameras have become fogged? You can't guarantee that all sensors will work all the time, so you need 2 or three of each one to ensure redundancy. This is what is required on airliners - 3 x hydraulic systems (even though the plane is flyable with one), 3 x Angle of Attack sensors etc.

 

[tess19]

I believe that FSD is a very good assistance to the driver. But an assistant is all it would be able to do. There are just too many situations for our car to be able to FSD. Call me a cynic but i don't believe the computer is learning All the time when all of us are on autopilot. Without someone telling the computer what is right and what is not, how can it tell what to do next time? The amount of information stored of every tesla drivers in the world would be enormous.

 

[Cogarch]

Without someone telling the computer what is right and what is not, how can it tell what to do next time? The amount of information stored of every tesla drivers in the world would be enormous.

In principle the system could note what the computer had been planning to do and compare it with what the driver actually did and so learn by reinforcement without further human input.

 

[MrBadger]

We used that on one of our test fleet trials - running shadow mode to validate AI choices vs driver choices whilst driving on national roads. The vehicles were not autonomous, but had autonomous driving systems running in them and hooked up to everything in the car. Even on small fleet, data sets were pretty big - iirc some of the original plans to transfer the data over public networks was dropped for being too slow and costly - so they build dedicated locations for the vehicles to park with private network access. I suspect it was 4G and WIFI so similar tech to what is every day in a Tesla.

MOVE UK

Tesla are way ahead of that game.

 

[Adopado]

In principle the system could note what the computer had been planning to do and compare it with what the driver actually did and so learn by reinforcement without further human input.

One issue is that I commonly drive on a relatively narrow "A" road that has quiet stretches with open corners that any sensible driver (me) takes a driving line that cuts well across the centre line. In those instances I can see the road quarter of a mile ahead is clear of oncoming traffic, but I am looking across open countryside to see that next section of road ... It's not just the road at the corner itself that I need to be clear. If my car is learning from the way I drive the road then I hope it doesn't think it can take that line in every circumstance!

 

[pgkevet]

I go one stage further on my country A roads and straddle the centre line on clear stretches to give a little more time if wildlife comes out of the hedge/verge or a tractor at a field turning as well as cutting corners where visibility is OK

 

[AndrewGR]

Actually, an airline pilot would probably make a half-decent hack of flying a glider, mainly because he/she will have flown light aircraft for hundreds of hours before graduating to large jet aircraft - and gliders handle much like light aircraft.

As a gliding instructor that occasionally has the opportunity to instruct airline (and military) pilots I can confidently assert that many, if not most, are incapable of making a ‘half-decent hack of flying a glider’ at least for their first few flights. Most particularly, they don’t realise that they have to use their feet (rudder) as well as the stick (ailerons and elevators) in order to ‘go round corners’. Many experienced light aircraft pilots have difficulties too. Strangely, sailing experience seems more useful in that sailors have a good feel for what their craft is doing. I’m still waiting for delivery of my Tesla but I’m fast realising that AP won’t be much use for most of the driving I do - which is not on Motorways or even dual carriageways. On rural roads the Ned to take over in a split second whilst going round bends sounds much too challenging to risk!

 

[gangzoom]

Without someone telling the computer what is right and what is not, how can it tell what to do next time? The amount of information stored of every tesla drivers in the world would be enormous.

Surely that's the point of AI, it will decided by itself and not need to resort to programming?

 

[SirRob]

For goodness sakes don't do that. She absolutely shouldn't be driving while snorting coke, you have to be able to have BOTH hands on the wheel. Completely irresponsible, don't do that again.

Don't tell me how to live my life!

 

[Cogarch]

Surely that's the point of AI, it will decided by itself and not need to resort to programming?

It depends on whether you are using supervised or unsupervised machine learning.

Supervised vs Unsupervised Learning: Key Differences

 

[pdk42]

As a gliding instructor that occasionally has the opportunity to instruct airline (and military) pilots I can confidently assert that many, if not most, are incapable of making a ‘half-decent hack of flying a glider’ at least for their first few flights. Most particularly, they don’t realise that they have to use their feet (rudder) as well as the stick (ailerons and elevators) in order to ‘go round corners’. Many experienced light aircraft pilots have difficulties too.

Well, as an ex gliding instructor I have to say that I don't agree with this. Maybe we've flown with a different breed of power pilots ! Sure, they'll be a long way from flying with any finesse, and getting turns balanced will probably elude them - but if their instructor passed out in the air, they'd probably get the thing down without killing themselves. Maybe that's setting the bar a bit low, but that's what I meant by a "half-decent hack".

 

[arg]

Call me a cynic but i don't believe the computer is learning All the time when all of us are on autopilot. Without someone telling the computer what is right and what is not, how can it tell what to do next time?

Tesla have claimed to be doing two things:

1. When driving on autopilot, if the driver intervenes they can capture the data leading up to the intervention and upload it for analysis.
2. They set up 'triggers' for whatever scenarios they are working on at the time - maybe "traffic lights with two lanes" or "roadworks", and whenever cars encounter these situations (even if not on autopilot at the time) they upload images, either for manual labelling prior to training the system for the first time, or to validate behaviour already 'finished'.

Exactly how much of this they do isn't disclosed - do they really process every autopilot intervention, or perhaps just locations where intervention occurs more than once?

There is also 'shadow mode' which they have often claimed to do, but have been less forthcoming on the exact details. Probably this means running two versions simultaneously - the old one actively driving the car, plus a new one under test - and comparing the results.

 

[adsheff]

One issue is that I commonly drive on a relatively narrow "A" road that has quiet stretches with open corners that any sensible driver (me) takes a driving line that cuts well across the centre line. In those instances I can see the road quarter of a mile ahead is clear of oncoming traffic, but I am looking across open countryside to see that next section of road ... It's not just the road at the corner itself that I need to be clear. If my car is learning from the way I drive the road then I hope it doesn't think it can take that line in every circumstance!

This is a very good point. On open roads you look way ahead for things. Even on a motorway, in heavy traffic I am looking way past the few cars in front for brake lights. If there is some queues forming a long way ahead, I'd ease off the gas. I've noticed AP will keep driving if the road immediately ahead is clear, and brake much later than I would. I realise this is how a lot of people drive anyway, but it it neither efficient nor safe, and is a big cause of traffic congestion - i.e. late braking and overly keen acceleration when it clears.

 

[elecTED]

Tesla are still stating the following on their website under the FSD purchase option:

Coming later this year:

• Recognise and respond to traffic lights and stop signs.
• Automatic driving on city streets.

Santa could be very busy this year...

My one-liner on Autopilot has not changed:

Autopilot feels like being controlled by an over-zealous driving instructor and behaves like an over-cautious learner driver.

 

[LukeT]

This is a very good point. On open roads you look way ahead for things. Even on a motorway, in heavy traffic I am looking way past the few cars in front for brake lights. If there is some queues forming a long way ahead, I'd ease off the gas. I've noticed AP will keep driving if the road immediately ahead is clear, and brake much later than I would. I realise this is how a lot of people drive anyway, but it it neither efficient nor safe, and is a big cause of traffic congestion - i.e. late braking and overly keen acceleration when it clears.

This is pretty much how I feel about it. If the car smoothed out the accelerations and decelerations compared with the human, rather than doing the opposite as it does, it would surely be more comfortable and more efficient. I think it would also be marginally safer, not because I think the AP car itself is going to crash either way, but because any car that is exacerbating congestion is increasing the chances of others having an accident. I'd have thought a modest adjustment this way would be possible over the air.

 

[adsheff]

This is pretty much how I feel about it. If the car smoothed out the accelerations and decelerations compared with the human, rather than doing the opposite as it does, it would surely be more comfortable and more efficient. I think it would also be marginally safer, not because I think the AP car itself is going to crash either way, but because any car that is exacerbating congestion is increasing the chances of others having an accident. I'd have thought a modest adjustment this way would be possible over the air.

Yep I'd love a super-eco mode. I'm quite happy taking 20 seconds or more to get up to 70mph, it makes no difference to your journey time but reduces power consumption massively.

 

[WannabeOwner]

If the car smoothed out the accelerations and decelerations compared with the human, rather than doing the opposite as it does, it would surely be more comfortable and more efficient

If it was my Gig and DEV team I would be concentrating on auto-driving features more than smooth efficient driving, and do that later.

FSD is a race between brands. Being first to drive coast-to-coast (or whatever target you have) being more important than doing it "well", at first release. Safely yes, efficiently / smoothly only "as best as you can without huge DEV effort"

 

[Big Earl]

Yep I'd love a super-eco mode. I'm quite happy taking 20 seconds or more to get up to 70mph, it makes no difference to your journey time but reduces power consumption massively.

I don’t see much difference in real world consumption between rapid and gradual acceleration. As long as you don’t over-accelerate followed by deceleration (speeding up to a red light), your overall efficiency should be similar whether you do 15 seconds 0-60 or 7 seconds. The motors are only slightly less efficient at moderate power output than they are at low power output.

One of my biggest pet peeves is when people dawdle away from the stop light, reducing intersection efficiency. If everyone got up to speed quickly, vehicle throughout per light cycle would be improved.

 

[Adopado]

One of my biggest pet peeves is when people dawdle away from the stop light, reducing intersection efficiency. If everyone got up to speed quickly, vehicle throughout per light cycle would be improved.

EVs excel at that intersection alacrity! The others are waiting for their auto-restart of the ICE to kick in, the mini turbo to spool up, sometimes for the auto gearbox to select a gear etc ... multiply that by a few cars ... and a reluctance to press down with the right foot ...

 

[WannabeOwner]

I'm quite happy taking 20 seconds or more to get up to 70mph, it makes no difference to your journey time but reduces power consumption massively.

The "effort" to get to 70 is much the same, i.e. getting all-that-weight from 0MPH to 70MPH ... some extra heat generated / energy wasted by putting your foot down in a P-model.

Once at that speed you have to push the air out of the way, so accelerating faster, and getting to 70 sooner, means air has to be pushed out of the way, at 70 MPH, for a slightly longer time. But not much difference between prompt "non-Ludicrous " acceleration and dawdling acceleration.