The Self-Driving Vehicles Thread

[slowsider]

You'll be wanting AI for that.

[Horse]

You'll be wanting AI for that.

But will it need to better than a human, and how would we know?

[slowsider]

A colourful solution?

https://www.theguardian.com/artanddesig ... dApp_Other

[Horse]

A colourful solution?

https://www.theguardian.com/artanddesig ... dApp_Other

A more permanent version:

[Horse]

Should AVs be equivalent to average drivers, or better? If better, how much better? How should this equivalence or superiority be assessed?

I'm glad you asked that, because it seems to me to be an important point to discuss.

But with the obvious 'what gets measured gets done' concerns.

It seemed likely that someone would have bothered.

https://www.researchgate.net/publicatio ... iving_cars

Makes some interesting points.

Although the level of safety required before drivers will accept self-driving cars is not clear, the criterion of being safer than a human driver has become pervasive in the discourse on vehicle automation. This criterion actually means "safer than the average human driver," because it is necessarily defined with respect to population-level data. At the level of individual risk assessment, a body of research has shown that most drivers perceive themselves to be safer than the average driver (the better-than-average effect). Method: Using an online sample, this study examined U.S. drivers' ratings of their own ability to drive safely and their desired level of safety for self-driving vehicles. Results: This study replicated the better-than average effect and showed that most drivers stated a desire for self-driving cars that are safer than their own perceived ability to drive safely before they would: (1) feel reasonably safe riding in a self-driving vehicle; (2) buy a self-driving vehicle, all other things (cost, etc.) being equal; and (3) allow self-driving vehicles on public roads. Conclusions: Since most drivers believe they are better than average drivers, the benchmark of achieving automation that is safer than a human driver (on average) may not represent acceptably safe performance of self-driving cars for most drivers. Practical applications: If perceived level of safety is an important contributor to acceptance of self-driving vehicles, the popular "safer than a human driver" benchmark may not be adequate for widespread acceptance.

Perhaps another consideration is what the vehicle is doing. We already have different standards for different vehicle licence categories, whether by age, tests or retraining - for types of vehicle which might be on the same stretch of road (eg moped, tractor car, hgv, bus, emergency services).

So, should there different 'tests' for AV systems, based on where they are used, what they carrying, etc?

There might even be other considerations which eventually need to be incorporated, for example the duty of care that an AV taxi might be expected to have if the only passenger is a child.

[Cousin Jack]

Good point!

As a non-fan of current AVs I would want to see MUCH higher standards imposed for eg, HGVs, PSVs, dangerous loads and high performance vehicles.

[slowsider]

Human drivers currently require additional qualifications for those. Bar the last.

[Mussels]

Human drivers currently require additional qualifications for those. Bar the last.

And the last is pointless if a robot is driving.

I'll just wait until I stop seeing stupid crashes in the press, then embrace the cheap chauffeur and head to the pub

[Horse]

I would want to see MUCH higher standards

Higher than *what*?

How would an 'acceptable' system be defined for safety, say for carrying coal or firewood, or 'higher' for carrying people, then 'wow!' for carrying a cub scout troop going to a fireworks party taking a shortcut through a hydrogen plant?

Assuming all the same (and equivalent to non-AVs) occupant and pedestrian protection), do you think it should be on 'performance' aspects? Hazard identification (quantity), reaction time, braking distances?

What else could it be graded on?

And would manufacturers deliberately downgrade / hold back tech improvements, apart from on a cost basis?

Human drivers currently require additional qualifications for those.

As, earlier, you were kind enough to point out that I hadn't taken full notice of one of your posts, I'll just remind you that:

different standards for different vehicle licence categories, whether by age, tests or retraining - for types of vehicle (eg moped, tractor car, hgv, bus, emergency services).

[Cousin Jack]

Higher than an 'average' driver, and preferably higher than the human test standard.

[Horse]

Higher than an 'average' driver, and preferably higher than the human test standard.

If you take the elements of a car driving tests:
- Highway code; should be ok, programme once it won't forget, then teach it to recognise signs, markings, etc (did you see the AV in the video break two laws?)
- Hazard perception; as in a video I posted some time ago, some systems are being soak tested on umpteen variations of actual crashes that have been studied in detail. Plus, the current HPT features computer animations! [See below]
- Manouevres; cars with self-park already exist. Auto Lane Keep, etc.
- General driving, negotiating hazards, etc. That could be a substantial 'test'.

What they probably can't do yet is the 'advanced' observations and insight, such as 'bits of straw on the road = farm tractor and trailer around the corner'. But that is, as in human driving, something that doesn't happen immediately in learning to drive. And, for some, never does.

[Cousin Jack]

But these little things, straw on the road, bins by the roadside, weather last night was heavy rain, are all the things that make most experienced drivers much safer than new drivers.

Conversely, some new drivers learn that 99 times out of 100 some things are safe. Then they do it for the 100th time and it all goes pete tong.

AI learning is fine, but how do you teach an AV to learn some things and not learn others?

[Horse]

But these little things, straw on the road, bins by the roadside, weather last night was heavy rain, are all the things that make most experienced drivers much safer than new drivers.

From a conversation a few years ago, with someone involved in test driving them, that's what they'd like to achieve.

So (and using very general descriptions), are we looking at an AV testing scale of CBT - DAS - IAM - police advanced?

Conversely, some new drivers learn that 99 times out of 100 some things are safe. Then they do it for the 100th time and it all goes pete tong.

Guessing, an AV would allow for that 1% to be anywhere from 1/100 to 100/100 events I don't think there's any suggestion of them electronically rolling dice and saying "what the hell" then accelerating

AI learning is fine, but how do you teach an AV to learn some things and not learn others?

Ground rules?
- keep between the kerbs and hedges
- don't hit anything
etc?

[Horse]

V2V comms would sort that. The most challenging period is the transition before all vehicles are equipped and you have to deal with hoomans.

Quite. There are vehicles on the road today that date back to the 1920s, so 100 years of chaos during the 'transition'.

In the news today:

https://highways-news.com/connected-vet ... athfinder/

Three veteran cars built at the turn of the 20th century have travelled from London to Brighton while using the latest in connected technology in what is believed to be the world’s first connected veteran car event.

The 1903 Daimler, 1901 Mors and 1900 Daimler were among the 350 cars which celebrated the 125th anniversary of the first Emancipation Run, by travelling the 60 miles from Hyde Park in London to Brighton’s seafront on Sunday 7 November. They were celebrating the passing of the 1896 Light Locomotives on Highways Act, which raised the national speed limit to 14mph and effectively allowed motor vehicles to develop as a practical means of transport.

The vehicles were fitted with new apps that integrate several services in one place – sharing their location and status with organisers, receiving navigation specific to the official route and also in-vehicle route information such as roadworks, congestion and safety messages. They therefore became the world’s oldest connected cars.

"Fitted with apps" actually means 'smartphines'.

[slowsider]

smartphines ??

Just to show I read your posts to the bitter end

[Horse]

smartphines ??

Just to show I read your posts to the bitter end

It's my accent.

[Horse]

Researchers in Australia say they have developed a technology that allows autonomous vehicles to track moving pedestrians hidden behind buildings and cyclists obscured by cars, trucks and buses.

The technology has been refined as part of a project funded by the iMove Cooperative Research Centre in collaboration with the University of Sydney’s Australian Centre for Field Robotics and Australian connected vehicle company Cohda Wireless. iMove recently released its new findings in a final report following three years of research and development.

The approach relies on collective perception (CP) (also known as cooperative perception). Using roadside ITS (Intelligent Transport System) information sharing units (‘ITS stations’), vehicles can share what they ‘see’ with others using vehicle-to-X (V2X) communication. The system significantly increases the vehicle’s range of perception by allowing it to tap into various viewpoints.

“This is a game changer for both human-operated and autonomous vehicles which we hope will substantially improve the efficiency and safety of road transportation,” said Professor Eduardo Nebot from the Australian Centre for Field Robotics.

“Using collective perception, the connected vehicle was able to track a pedestrian visually obstructed by a building. This was achieved seconds before its local perception sensors or the driver could possibly have seen the same pedestrian around the corner, providing extra time for the driver or the navigation stack to react to this safety hazard,” he added.

Another experiment demonstrated how collective perception could allow vehicles to safely interact with walking pedestrians, with the vehicle’s response based on the perception information provided by a roadside ITS station.

The three-year project also demonstrated the expected behavior of a connected vehicle when interacting with a pedestrian rushing towards a designated crossing area.

“Using the ITS system, the connected autonomous vehicle managed to take pre-emptive action: braking and stopping before the pedestrian crossing area based on the predicted movement of the pedestrian,” Nebot noted.

“The pedestrian tracking, prediction, path planning and decision making were based on the perception information received from the ITS roadside stations. Collective perception enables the smart vehicles to break the physical and practical limitations of onboard perception sensors.”

https://www.autonomousvehicleinternatio ... rners.html

[slowsider]

Be interesting to see how it helps human-operated vehicles. Theres a data-processing limit to the OS with them

[Horse]

Be interesting to see how it helps human-operated vehicles. Theres a data-processing limit to the OS with them

Yesterday evening watched one of the BBC's Crash Detectives programmes.

One fatal collision was where someone walked in front of an HGV at traffic lights. She timed it (shown by video recording) just as the lights went green.

Some form of detector on the vehicle, or additional warning, might have help prevent it.

[Horse]

Be interesting to see how it helps human-operated vehicles. Theres a data-processing limit to the OS with them

A mate has a BMW car with full 'surround' video coverage and the reversing system (video with 2x sets of lines and bleeps) also has a '3D' overlay.

Not too much of a stretch to expand it (in theory).

But attention on one thing is distraction from another.