self driving vehicle on UK motorways in 2023

[Count Steer]

Humans learn from imagination.

I suspect AI only learns from collisions, or at best, from very near misses.

The consistently repeating three gotcha bike crash types (smidsy, corners, overtaking) suggest that human motorcyclists can't learn others's (or even their own) collisions and near misses.

Smidsy's an interesting one. Fit the car with an AI and it won't pull out in front of you (but, is the car fitted with AI and how would you know?). However, there are junctions that AI cars will never get out of. Fit the bike with one and every time it identifies a car waiting to pull out it'll slow down, incrementally as it gets closer*. (So the car driver will be inclined to pull out). Unless the AIs have an acceptable risk level programmed in.

I foresee lots of AI vehicles covered in minor scuffs and scrapes.

* Unless it can interrogate the car.
Bike: Have I got a network connection? Yup.
Bike: Hello car, are you AI?
Car: Incoming signal, protocol not recognised
Bike: Hello car, are you AI?
Car: Protocols update required, downloading.
Bike: No handshake, decreasing speed
Car: No 6G available, download paused
Car 2: (Following bike) Bike are you stopping?
Bike: Incoming signal, protocol not recognised, Apple 7.1.2.7.34.94b drivers installing

Ad nauseum.

[Horse]

Humans learn from imagination.

I suspect AI only learns from collisions, or at best, from very near misses.

The consistently repeating three gotcha bike crash types (smidsy, corners, overtaking) suggest that human motorcyclists can't learn others's (or even their own) collisions and near misses.

So the algorithms factor in severity of collision or potential severity of near miss and adjust accordingly. We all know where that leads don't we? Nothing travels at more than 10mph. BRAKE will be ecstatic.

Aren't motorways supposed to be the 'safest' roads?

Wouldn't even a 10mph collision be a tad uncomfortable for a pedestrian or cyclist.

[Horse]

Smidsy's an interesting one. Fit the car with an AI and it won't pull out in front of you (but, is the car fitted with AI and how would you know?).

Interesting alternative to "but I had my headlamp on - and he still pulled out!"

[Horse]

Imagine if humans had a wi-fi connection to each other and learned from mistakes made by those linked to them.

As an aside: I saw the 'HELM' convoy of three 'connected' HGVs heading down the M5. There was a fourth, non-convoy, truck tagging along - at about the same following 'distance'

[slowsider]

Imagine if humans had a wi-fi connection to each other and learned from mistakes made by those linked to them.

As an aside: I saw the 'HELM' convoy of three 'connected' HGVs heading down the M5. There was a fourth, non-convoy, truck tagging along - at about the same following 'distance'

Monkey see, monkey do?

[Count Steer]

Imagine if humans had a wi-fi connection to each other and learned from mistakes made by those linked to them.

As an aside: I saw the 'HELM' convoy of three 'connected' HGVs heading down the M5. There was a fourth, non-convoy, truck tagging along - at about the same following 'distance'

Haven't tried it (haven't even tried ordinary cruise control) but, according to the manual I can set the cruise control on the wife's car to 'follow that car/lorry'. Probably not at distance = 1m though.

It has blind spot sensors and can interfere with the steering or brakes in certain situations etc. Small step to 'follow, changing lanes as required'. Might even already be able to...

[The Spin Doctor]

Humans learn from imagination.

I suspect AI only learns from collisions, or at best, from very near misses.

The consistently repeating three gotcha bike crash types (smidsy, corners, overtaking) suggest that human motorcyclists can't learn others's (or even their own) collisions and near misses.

To a certain extent that's true...

...but it also highlights the real weakness of rider education and road safety generally. We're not taught 'not to crash'. We're taught to 'drive properly' in the mistaken assumption that will stop people crashing.

Other industries have changed. Flying 'right' didn't work 80 years ago when teaching pilots to fly complex planes in WW2 so they changed their approach - firstly to recovering from emergencies and then to understanding how a particular problem might cause an emergency and how to deal with it to prevent the emergency developing.

That's a simplification of course, but it highlights the difference in philosophy. The blame for crashing doesn't entirely lie with the humans who get it wrong. And that's why I've spent the last decade developing Science Of Being Seen campaign that seeks to forewarn riders of the rears and circumstances in which SMIDSYs happen, and the No Surprise? No Accident! approach to riding.

[The Spin Doctor]

Smidsy's an interesting one. Fit the car with an AI and it won't pull out in front of you (but, is the car fitted with AI and how would you know?). However, there are junctions that AI cars will never get out of. Fit the bike with one and every time it identifies a car waiting to pull out it'll slow down, incrementally as it gets closer*. (So the car driver will be inclined to pull out). Unless the AIs have an acceptable risk level programmed in.

There was a crash a few years ago - walking pace, so no serious damage - involving a Googlecar in their very limited trial area and a bus. The Googlecar was in the inside lane which was blocked by road works. The car started to edge out into the adjacent lane, 'expecting' a vehicle to let it out. The bus driver didn't. The car kept 'edging' right up to the moment it hit the bus.

The programming simply hadn't taken into account the fact that the vehicle might meet a git behind the wheel who'd deliberately block the Googlecar in.

The 'slow down until you are travelling infinitely slowly and can stop in an infinitely short distance' paradox is one I've talked about before. Unless we ride very slowly and keep a very wide gap, at some point we MUST commit ourselves into the 'killing zone' where we cannot avoid a collision if the vehicle emerges at just the wrong moment. We can SHRINK that zone by changing speed and position, but on many roads it's impossible to eliminate it completely.

[Count Steer]

Smidsy's an interesting one. Fit the car with an AI and it won't pull out in front of you (but, is the car fitted with AI and how would you know?). However, there are junctions that AI cars will never get out of. Fit the bike with one and every time it identifies a car waiting to pull out it'll slow down, incrementally as it gets closer*. (So the car driver will be inclined to pull out). Unless the AIs have an acceptable risk level programmed in.

There was a crash a few years ago - walking pace, so no serious damage - involving a Googlecar in their very limited trial area and a bus. The Googlecar was in the inside lane which was blocked by road works. The car started to edge out into the adjacent lane, 'expecting' a vehicle to let it out. The bus driver didn't. The car kept 'edging' right up to the moment it hit the bus.

The programming simply hadn't taken into account the fact that the vehicle might meet a git behind the wheel who'd deliberately block the Googlecar in.

The 'slow down until you are travelling infinitely slowly and can stop in an infinitely short distance' paradox is one I've talked about before. Unless we ride very slowly and keep a very wide gap, at some point we MUST commit ourselves into the 'killing zone' where we cannot avoid a collision if the vehicle emerges at just the wrong moment. We can SHRINK that zone by changing speed and position, but on many roads it's impossible to eliminate it completely.

We could commit if we've made eye contact with the driver. Love to see them build that into a production car AI.

[slowsider]

Smidsy's an interesting one. Fit the car with an AI and it won't pull out in front of you (but, is the car fitted with AI and how would you know?). However, there are junctions that AI cars will never get out of. Fit the bike with one and every time it identifies a car waiting to pull out it'll slow down, incrementally as it gets closer*. (So the car driver will be inclined to pull out). Unless the AIs have an acceptable risk level programmed in.

There was a crash a few years ago - walking pace, so no serious damage - involving a Googlecar in their very limited trial area and a bus. The Googlecar was in the inside lane which was blocked by road works. The car started to edge out into the adjacent lane, 'expecting' a vehicle to let it out. The bus driver didn't. The car kept 'edging' right up to the moment it hit the bus.

The programming simply hadn't taken into account the fact that the vehicle might meet a git behind the wheel who'd deliberately block the Googlecar in.

The 'slow down until you are travelling infinitely slowly and can stop in an infinitely short distance' paradox is one I've talked about before. Unless we ride very slowly and keep a very wide gap, at some point we MUST commit ourselves into the 'killing zone' where we cannot avoid a collision if the vehicle emerges at just the wrong moment. We can SHRINK that zone by changing speed and position, but on many roads it's impossible to eliminate it completely.

We could commit if we've made eye contact with the driver. Love to see them build that into a production car AI.

[The Spin Doctor]

We could commit if we've made eye contact with the driver. Love to see them build that into a production car AI.

One of the things I explain on Science Of Being Seen is that 'eye contact' is a very iffy concept. With the human foveal zone being 5 degrees across someone can appear to be looking straight at you, and actually be focused on the vehicle behind. How many riders have you heard say "the driver was looking right at me and still pulled out"... that's why.

Watch the number of times the driver's head suddenly snaps round to focus on you when you're right on top of the vehicle - that's the moment they have actually seen the bike and it's often scarily close.

The only thing that we can say with any certainty is that if the driver's head is turned the other way, they can't see us.

[Count Steer]

We could commit if we've made eye contact with the driver. Love to see them build that into a production car AI.

One of the things I explain on Science Of Being Seen is that 'eye contact' is a very iffy concept. With the human foveal zone being 5 degrees across someone can appear to be looking straight at you, and actually be focused on the vehicle behind. How many riders have you heard say "the driver was looking right at me and still pulled out"... that's why.

Watch the number of times the driver's head suddenly snaps round to focus on you when you're right on top of the vehicle - that's the moment they have actually seen the bike and it's often scarily close.

The only thing that we can say with any certainty is that if the driver's head is turned the other way, they can't see us.

I hesitated before using the term. I really meant the driver has actually, at some point, turned their head in this general direction. In the future you'll probably be watching for either a) a look of horror on the 'drivers' face as the autonomous vehicle decides to go in a situation that they wouldn't or b) the beard the (male) 'driver' has grown waiting for the AI to decide it's safe to proceed.

[Horse]

Unless we ride very slowly and keep a very wide gap, at some point we MUST commit ourselves into the 'killing zone' where we cannot avoid a collision if the vehicle emerges at just the wrong moment. We can SHRINK that zone by changing speed and position, but on many roads it's impossible to eliminate it completely.

Although on motorways everyone is usually travelling the same direction

I saw a presentation several years ago about AVs taking risks. The example given was an AV in lane 1 with another vehicle intending to merge from a slip road. The AV could change lanes to allow room, but would need to accept shorter headroom in lane 2.


Since I haven't heard any stories of all AVs coming to a halt on a major road where there is a vehicle waiting to emerge, I presume it doesn't actually happen, so Brake will be disappointed.

[Count Steer]

Unless we ride very slowly and keep a very wide gap, at some point we MUST commit ourselves into the 'killing zone' where we cannot avoid a collision if the vehicle emerges at just the wrong moment. We can SHRINK that zone by changing speed and position, but on many roads it's impossible to eliminate it completely.

Although on motorways everyone is usually travelling the same direction

I saw a presentation several years ago about AVs taking risks. The example given was an AV in lane 1 with another vehicle intending to merge from a slip road. The AV could change lanes to allow room, but would need to accept shorter headroom in lane 2.


Since I haven't heard any stories of all AVs coming to a halt on a major road where there is a vehicle waiting to emerge, I presume it doesn't actually happen, so Brake will be disappointed.

Not enough of them around yet Hoss and it will depend how (if) they learn. Get enough smidsy type situations under their belt...what will they do? Adjust a parameter, refine a probability calculation, hand control over to the driver (the 'all yours mate' subroutine), read the Science Of Being Seen, book themselves in for fitting with better brakes and a new wally vest? They'll probably really just keep getting smidsy'd....or go slower.

If they're that good, they're going to have to be unrecognisable...or everyone will go 'ah, computer on wheels, I'll pull out, it won't hit me'.

Really, it's only code and most of what's touted as AI is just rules based stuff anyway. It's one thing having a few lorries in convoy on a near perfect road but I don't think London cabbies have much to fear yet (apart from Uber ).

[Cousin Jack]

Humans learn from imagination.

I suspect AI only learns from collisions, or at best, from very near misses.

The consistently repeating three gotcha bike crash types (smidsy, corners, overtaking) suggest that human motorcyclists can't learn others's (or even their own) collisions and near misses.

[Horse]

The only thing that we can say with any certainty is that if the driver's head is turned the other way, they can't see us.

MSF used to distinguish between looking, seeing and understanding.

[Horse]

Since I haven't heard any stories of all AVs coming to a halt on a major road where there is a vehicle waiting to emerge, I presume it doesn't actually happen, so Brake will be disappointed.

Not enough of them around yet Hoss and it will depend how (if) they learn.

I'm sure that if it happened at all often, word would get around (like the one in the US that couldn't cope with roadworks, so stopped.

Really, it's only code and most of what's touted as AI is just rules based stuff anyway.

Well over two years ago (very early lockdown 1), I saw a short video recorded from a car being driven through London (clip ended in The Mall).

The processing system added 'bounding boxes' around everything it identified. Even to the extent of 'horse' and 'rider' being identified separately!

Then layer them rules on top and there's potential for a good prediction system.

[Horse]

FWIW, several AV developers are working on camera-based systems. Obvious concern there on glare etc.

This was added to YouTube in 2019, so I guess it's more impressive now?

[Count Steer]

Since I haven't heard any stories of all AVs coming to a halt on a major road where there is a vehicle waiting to emerge, I presume it doesn't actually happen, so Brake will be disappointed.

Not enough of them around yet Hoss and it will depend how (if) they learn.

I'm sure that if it happened at all often, word would get around (like the one in the US that couldn't cope with roadworks, so stopped.

Really, it's only code and most of what's touted as AI is just rules based stuff anyway.

Well over two years ago (very early lockdown 1), I saw a short video recorded from a car being driven through London (clip ended in The Mall).

The processing system added 'bounding boxes' around everything it identified. Even to the extent of 'horse' and 'rider' being identified separately!

Then layer them rules on top and there's potential for a good prediction system.

Oh, I'm sure there are powerful prototypes out there and computing power does get cheaper. If the cost benefit makes sense (To who though? Who actually wants this technology?) some countries will gradually allow/enforce more and more autonomy/constraint options. We're a long, long way off the way some of it's being pitched though - we're getting all the sales videos, some of which are aimed at getting funding, not putting stuff in a Fiesta.

It's like a clash of dreams and reality. eg next thing up is hindering exceeding the speed limits. The % of the time that our latest acquisition doesn't know the correct limit is (Overgrown/missing signs) the GPS is better but is eg completely unaware of the 10mph limit in a big stretch of roadworks. The lane centring thing got turned off...have you seen the state of road markings? I'm sure it works fine on a beautifully painted new road, with good lighting at night etc etc. TBH even the relatively simple stuff they've managed to get in run-of-the-mill production cars is a bit pants in the real world.

By the time it all works half properly I'll be past caring, but I'll have to suffer the 'half-baked' phase. I still predict automation will slow everything on the roads down for quite a few years, which will be novel as it usually speeds things up. (What was the average speed of that car in London, during lockdown?).

[Horse]

I think I've said before, personal opinion is that 'cars' is a way of developing technology in a hardest case environment.

Actual applications are likely to be extensively elsewhere.

eg:
https://www.oxbotica.com/our-applications/