The Lemon Fool

Howard wrote:
Ody. Can you justify your statement? Source?

Sadly Tesla doesn’t agree with you. It seems to agree with my experience actually driving the car.

The very latest car’s spec has cameras which Tesla state have a maximum range of 250 metres.

Even the ultrasonic sensors have a range of less than 500 metres.

It’s fun to dream but do facts get in the way?

Worth reading what Tesla claim for autopilot for their US cars in 2022.

regards

Howard

"Eight surround cameras provide 360 degrees of visibility around the car at up to 250 meters of range. Twelve updated ultrasonic sensors complement this vision, allowing for detection of both hard and soft objects at nearly twice the distance of the prior system."

https://www.tesla.com/autopilot

The ultrasonics have a range of 8 metres. Presumably twice the range refers to a change in the camera array.

BobbyD wrote:

Howard wrote:
Ody. Can you justify your statement? Source?

Sadly Tesla doesn’t agree with you. It seems to agree with my experience actually driving the car.

The very latest car’s spec has cameras which Tesla state have a maximum range of 250 metres.

Even the ultrasonic sensors have a range of less than 500 metres.

It’s fun to dream but do facts get in the way?

Worth reading what Tesla claim for autopilot for their US cars in 2022.

regards

Howard

"Eight surround cameras provide 360 degrees of visibility around the car at up to 250 meters of range. Twelve updated ultrasonic sensors complement this vision, allowing for detection of both hard and soft objects at nearly twice the distance of the prior system."

https://www.tesla.com/autopilot

The ultrasonics have a range of 8 metres. Presumably twice the range refers to a change in the camera array.

Oh! Yes, you are right .

So Tesla are confirming that the latest spec car can't "see" as far a human. In fact less than 250 metres. Which I, and others have found is a major problem in practice driving at motorway speeds approaching an obstacle.

Ody's FSD dream seems a long way off still?

Too far ahead for a camera to make out?

regards

Howard

Howard wrote:

BobbyD wrote:

Howard wrote:
Ody. Can you justify your statement? Source?

Sadly Tesla doesn’t agree with you. It seems to agree with my experience actually driving the car.

The very latest car’s spec has cameras which Tesla state have a maximum range of 250 metres.

Even the ultrasonic sensors have a range of less than 500 metres.

It’s fun to dream but do facts get in the way?

Worth reading what Tesla claim for autopilot for their US cars in 2022.

regards

Howard

"Eight surround cameras provide 360 degrees of visibility around the car at up to 250 meters of range. Twelve updated ultrasonic sensors complement this vision, allowing for detection of both hard and soft objects at nearly twice the distance of the prior system."

https://www.tesla.com/autopilot

The ultrasonics have a range of 8 metres. Presumably twice the range refers to a change in the camera array.

Oh! Yes, you are right .

So Tesla are confirming that the latest spec car can't "see" as far a human. In fact less than 250 metres. Which I, and others have found is a major problem in practice driving at motorway speeds approaching an obstacle.

Ody's FSD dream seems a long way off still?

Too far ahead for a camera to make out?

regards

Howard

70 miles per hour is 31.3 metres per second.

So 250 yards is 7.99 seconds which is a very long time for a computer.

However, the 250 m is for 360 degree vision, the linear distance will be a lot further.

Regards,

mc2fool wrote:

ursaminortaur wrote:

murraypaul wrote:

odysseus2000 wrote:If the testing routine is good it should catch any serious errors before the neural net is sent to the wild.

Yes, that is what should happen.

It isn't. Tesla is releasing beta software out to the wild and letting the real world be the test.

Which at some point makes sense as then the cars encounter real world situations rather than just the artificial situations dreamt up by your testers.

As an ex-software engineer I can tell y'all that it is axiomatic that no in-house testing procedures will catch all the bugs that real users in the real world will uncover and suffer. That's what beta testing is about (and even then...)

Yes, but computers fly & land rockets & jet airliners.

Sure there may be bugs, but these systems have for years been good enough.

Regards,

odysseus2000 wrote:

mc2fool wrote:

ursaminortaur wrote:

murraypaul wrote:

odysseus2000 wrote:If the testing routine is good it should catch any serious errors before the neural net is sent to the wild.

Yes, that is what should happen.

It isn't. Tesla is releasing beta software out to the wild and letting the real world be the test.

Which at some point makes sense as then the cars encounter real world situations rather than just the artificial situations dreamt up by your testers.

As an ex-software engineer I can tell y'all that it is axiomatic that no in-house testing procedures will catch all the bugs that real users in the real world will uncover and suffer. That's what beta testing is about (and even then...)

Yes, but computers fly & land rockets & jet airliners.

Sure there may be bugs, but these systems have for years been good enough.

Regards,

Yes, and you can be sure that they all got beta tested and did not rely on just an in-house testing routine.

Of course, even that doesn't always catch serious problems, ask the 737 MAX victims' families...

odysseus2000 wrote:Yes, but computers fly & land rockets & jet airliners.

Sure there may be bugs, but these systems have for years been good enough.

Airspace is much more tightly controlled than roadspace. Vertical and horizontal separation between aircraft is vastly greater than that for road vehicles. Just watch the aircraft on the approach path. Compare that with cars approaching a traffic signal.

TJH

tjh290633 wrote:

odysseus2000 wrote:Yes, but computers fly & land rockets & jet airliners.

Sure there may be bugs, but these systems have for years been good enough.

Airspace is much more tightly controlled than roadspace. Vertical and horizontal separation between aircraft is vastly greater than that for road vehicles. Just watch the aircraft on the approach path. Compare that with cars approaching a traffic signal.

TJH

Yes, and without neural nets robotic driving was never going to be practical, but neural nets are here & offer the potential to reduce road traffic deaths & serious injuries from the current 10 per day to some smaller number.

We are now in the “can it be done” phase moving towards the point where regulators have to decide if it is good enough.

Regards,

odysseus2000 wrote:neural nets are here & offer the potential to reduce road traffic deaths & serious injuries from the current 10 per day to some smaller number.

Can you flesh this out for me?
Explain what you understand happens and how it is done?

-sd

servodude wrote:

odysseus2000 wrote:neural nets are here & offer the potential to reduce road traffic deaths & serious injuries from the current 10 per day to some smaller number.

Can you flesh this out for me?
Explain what you understand happens and how it is done?

-sd

No

There are numerous videos and other sources of information on the internet about this technology presented by people who know much more about it than I do.

Regards,

odysseus2000 wrote:

servodude wrote:

odysseus2000 wrote:neural nets are here & offer the potential to reduce road traffic deaths & serious injuries from the current 10 per day to some smaller number.

Can you flesh this out for me?
Explain what you understand happens and how it is done?

-sd

No

There are numerous videos and other sources of information on the internet about this technology presented by people who know much more about it than I do.

Regards,

No worries - just thought you might have had some insider info given your certainty

Stuff round this moves really quickly and I've seen really impressive results recently in other domains (e.g. EEG analysis & sleep scoring)
- but even there, where a cockup isn't going to be immediately fatal, there is the perception that human corroboration is going to be required for a good period of time

- sd

odysseus2000 wrote:Yes, but in the UK there are 3000 deaths & serious injuries on the roads. You think a politician will stand in the way of reducing this carnage & financial cost?

And the clue here is railways.
Driverless trains have been available for years. Since train driving relies very little on looking out of the window, and there is (should be) no interaction with pedestrians, it is easy to replace train drivers with machines.
Errors by manual drivers are still responsible for more than half of rail accidents.
A fully automated system would actually be cheaper than the half-auto-half-manual system the railways are currently installing network-wide at vast expense.
And have we made any progress in getting rid of train drivers? - No, virtually none. There are 2 tube lines and the docklands where they do nothing but push a "go" button. On the former 2 they're still paid to sit there. Other than that, nowt.

Gryff

tjh290633 wrote:

odysseus2000 wrote:Yes, but computers fly & land rockets & jet airliners.

Sure there may be bugs, but these systems have for years been good enough.

Airspace is much more tightly controlled than roadspace. Vertical and horizontal separation between aircraft is vastly greater than that for road vehicles. Just watch the aircraft on the approach path. Compare that with cars approaching a traffic signal.

TJH

...and autopilots are operated in a highly controlled multi-operator environment designed specifically to minimise and eradicate points of error or confusion by trained operators on a continual path of training and assessment informed by forensic analysis of past failures, not some tech bro who wants to up his mindfulness app rating on the commute in to the office.

gryffron wrote:

odysseus2000 wrote:Yes, but in the UK there are 3000 deaths & serious injuries on the roads. You think a politician will stand in the way of reducing this carnage & financial cost?

And the clue here is railways.
Driverless trains have been available for years. Since train driving relies very little on looking out of the window, and there is (should be) no interaction with pedestrians, it is easy to replace train drivers with machines.
Errors by manual drivers are still responsible for more than half of rail accidents.
A fully automated system would actually be cheaper than the half-auto-half-manual system the railways are currently installing network-wide at vast expense.
And have we made any progress in getting rid of train drivers? - No, virtually none. There are 2 tube lines and the docklands where they do nothing but push a "go" button. On the former 2 they're still paid to sit there. Other than that, nowt.

Gryff

Yes, this is correct, but we are moving into a huge shift in the economy. As things now look to me we will see machines replace many currently well paid people & that shift will move throughout the economy. Suggesting that because things have been done one way in the past is no guarantee that they will continue that way. Once neural nets are seen to work very large numbers of jobs will cease to use humans giving much lower cost, better safety, but with a level of social disruption like we have never seen before.

Regards,

BobbyD wrote:

tjh290633 wrote:

odysseus2000 wrote:Yes, but computers fly & land rockets & jet airliners.

Sure there may be bugs, but these systems have for years been good enough.

Airspace is much more tightly controlled than roadspace. Vertical and horizontal separation between aircraft is vastly greater than that for road vehicles. Just watch the aircraft on the approach path. Compare that with cars approaching a traffic signal.

TJH

...and autopilots are operated in a highly controlled multi-operator environment designed specifically to minimise and eradicate points of error or confusion by trained operators on a continual path of training and assessment informed by forensic analysis of past failures, not some tech bro who wants to up his mindfulness app rating on the commute in to the office.

And it should also be noted that fully automatic takeoff and landing systems for commercial aircraft are almost non-existent so even if the pilots are not really needed for the routine flying part they are still needed at the start and end of flights.

ursaminortaur wrote:And it should also be noted that fully automatic takeoff and landing systems for commercial aircraft are almost non-existent so even if the pilots are not really needed for the routine flying part they are still needed at the start and end of flights.

I don't know. You could let a Chipmunk take itself off once lined up, without touching the controls. Done that a few times. If the tailwheel lifted off the ground, you could get some precession, but done properly the attitude did not change.

And what about Autoland on the Tridents? That was done many a time.

TJH

tjh290633 wrote:

ursaminortaur wrote:And it should also be noted that fully automatic takeoff and landing systems for commercial aircraft are almost non-existent so even if the pilots are not really needed for the routine flying part they are still needed at the start and end of flights.

I don't know. You could let a Chipmunk take itself off once lined up, without touching the controls. Done that a few times. If the tailwheel lifted off the ground, you could get some precession, but done properly the attitude did not change.

And what about Autoland on the Tridents? That was done many a time.

TJH

Note I said fully automated. Autoland requires pilot training in the system, pilot configuration and monitoring and hence isn't fully automated. And although small private aircraft may be able to take off in the manner you describe no commercial aircraft takes of automatically.


https://www.flightdeckfriend.com/ask-a-pilot/can-a-plane-land-automatically

Automatic landings probably account for less then 1% of all landings on commercial flights. Many pilots actually think it’s much easier to land the aircraft manually, as monitoring the auto-pilot in the autoland stage of flight is itself very demanding with a very high level of vigilance required at all stages.
.
.
.
Automatic landings require a high level of automation monitoring that needs retraining every six months for professional pilots. Autolands can only be performed under strict conditions that require the certification of both the aircraft (often downgraded to no autoland capability due to technical issues), both of the pilots and the airport itself. The pilots are still required to configure the aircraft and control it’s speed and monitor its flight path. Any number of relatively minor technical issues can compromise many fail-passive auto land systems, requiring a missed approach to be carried out and then a possible diversion to an airport which is clear of fog or low cloud.

An Autoland is often referred to as a CAT III (3) approach. This refers to the category of the Instrument Landing System (ILS), which is a radio aid used to direct the pilots towards the runway on the final approach stages of flight. Not all airports have a CAT III ‘runway’ which means that not all airports can support Autoland operations. Equally, autoland capability is not a function that all aircraft have.
.
.
.
Can a plane take-off automatically?

No. Commercial passenger jets are not able to take-off automatically. Currently no commercial aircraft has an auto take-off capability.

To dispel the myth; the vast majority of commercial aircraft (including all Boeings and Airbus’) have no automatic take-off capability – all take-offs must be completed manually by the pilots with the auto-pilot usually engaged at around 1,000 ft above the ground.

Tesla and others running self-driving tests collect all their sensor input and report back interesting cases to HQ. So any new version of the software is tested against a cumulative experience of millions of road miles. They also note where the human driver actually in charge currently deviated from the computer's prediction to refine their model.

So the AI not only has driven further than any human, it learns better from its mistakes. The current edge for a human is a better understanding of the real world and predictive capacity as to how other humans will act, but the ability of AI systems to label features and forecast their intentions is catching up fast.

The barriers to adoption are likely to be social, as the bar the AI will need to pass will be so much higher than a human driving test, and legal, because it would be so tempting to sue a big corporation for any weaknesses. Then there is the matter of psychology and perception of risk, you might accept a lift from Dave even though he's a bad driver, but not from an AI however impeccable their record.

ursaminortaur wrote:
An Autoland is often referred to as a CAT III (3) approach. This refers to the category of the Instrument Landing System (ILS), which is a radio aid used to direct the pilots towards the runway on the final approach stages of flight. Not all airports have a CAT III ‘runway’ which means that not all airports can support Autoland operations. Equally, autoland capability is not a function that all aircraft have.

Of course they need a suitable ILS system to perform. And no, not all aircraft have Autoland capability.

'Twas ever thus.

TJH

JohnB wrote:Tesla and others running self-driving tests collect all their sensor input and report back interesting cases to HQ. So any new version of the software is tested against a cumulative experience of millions of road miles. They also note where the human driver actually in charge currently deviated from the computer's prediction to refine their model.

So the AI not only has driven further than any human, it learns better from its mistakes. The current edge for a human is a better understanding of the real world and predictive capacity as to how other humans will act, but the ability of AI systems to label features and forecast their intentions is catching up fast.

The barriers to adoption are likely to be social, as the bar the AI will need to pass will be so much higher than a human driving test, and legal, because it would be so tempting to sue a big corporation for any weaknesses. Then there is the matter of psychology and perception of risk, you might accept a lift from Dave even though he's a bad driver, but not from an AI however impeccable their record.

I can think of human drivers I prefer to avoid.

Once AI appears and shows significantly lower accident rates, the transition to it will imho be rapid after a slowish start.

If one recalls the transition to on-line banking, it was initially slow and the media was full of how folk had been scammed, but once the advantages became known, the scam rate determined to be very low with guarantees about folk being scammed, large numbers of people moved over to online for most of their financial transactions and have not gone back to traditional methods. Sure there are some folk who still don't trust the internet for anything, but they are becoming an ever smaller component of the economy, ditto mobile phones.

Regards,

odysseus2000 wrote:
Once AI appears and shows significantly lower accident rates, the transition to it will imho be rapid after a slowish start.

If one recalls the transition to on-line banking, it was initially slow and the media was full of how folk had been scammed, but once the advantages became known, the scam rate determined to be very low with guarantees about folk being scammed, large numbers of people moved over to online for most of their financial transactions and have not gone back to traditional methods. Sure there are some folk who still don't trust the internet for anything, but they are becoming an ever smaller component of the economy, ditto mobile phones.

Regards,

I didn't realise online banking was an AI exercise and humans weren't involved in setting up payments, checking balances, transferring between accounts etc.

How fast is that transitioning of driverless trains coming on? I think you are underestimating both times with respect to that "rapid" transition and "slowish start"