Holy Crap I almost got T-Boned an hour ago. Middle lane at green light, wet road. Instinctively let car to my left go into intersection first. Car I couldn't see blasts thru intersection against a solid red. If I had gone faster out of the gate I would be totaled right now into my drivers side door.

@SwiftOnSecurity This is the embodiment of one of the core things we teach in advanced driving. Seeing the accident pinch points in time and not being there. Observation and understanding of your surroundings are key.

For folk who don't have extensive experience or depth of understanding, watch dash-cam vids on youtube, seriously (preferably from your own locale ― local driving habits can vary significantly). You get the chance to study and understand the anatomy of the accident. See how it unfolds and what led up to it. Learn from their mistakes. It's free and there are literally millions of them.

@McCovican @SwiftOnSecurity
When I worked on autonomous vehicles, we had a whole test dataset transcribed from youtube dashcam videos.
Between those and the NHTSA|FARS reports, you can build a really good intuition of how fatal/photogenic crashes move and develop.
@nonnihil tell me more?
@SwiftOnSecurity Well, I've been out of that line of work for a while, but as of 2016ish:
Almost no crash is one thing, but if you focus on the first detectable deviation from normality, it's nearly always a lane divergence or a late/missing braking.
Which means that the ability to detect the first hint of a change in acceleration or curvature is the whole ballgame.
(Well, that and good lane divergence sensors. If you invest in only one piece of modern auto ADAS tech, that one.)
@SwiftOnSecurity There's [EDIT There was a wrong number here previously] 12 fatalities per billion miles of travel. IIRC about a third of those are intoxication; about a fifth are too unique to prevent. The remaining 40%, about half lane divergence, the other half failing to stop/slow.
Of those, you could usually get a solid third of a second over human reaction times with a fast detection pipeline. That's forever in steering time, but much less braking energy, so steer-to-avoid was the focus.
@SwiftOnSecurity
Particularly on cars with a steering motor (modern power steering, older VGRS systems), an autonomous system can steer _much_ faster than a human without breaking traction, whereas braking is hard limited by physics. Braking works better at low speeds, steering at high speeds, and we were looking at fatal crashes specifically which are more often high speed.
And that also means closed-course testing with fancy racing drivers and inflatable obstacles, which is never not fun.
@SwiftOnSecurity So much fun we had to hand the project back to our parent company, as it turned out.
Now I just give robots knives; much safer.

@nonnihil @SwiftOnSecurity for selfish reasons I'm wondering if the equation of steering vs breaking changes in slippery conditions? Both in terms of responsiveness and risk of unintentional loss of control.

I'm thinking ice, and thick slush in particular.

@beeoproblem @SwiftOnSecurity
Huge problem! Esp. as _certain folks_ test their vehicles in CA and AZ with little weather.
ADAS systems have to estimate traction. Fortunately we have super high-rate telemetry on wheel angular rate versus inertials versus axle loading so detecting when traction starts to break is easy.
(Or do it with a mechanical gizmo -- old anti-lock brakes were purely mechanical)
There are also people doing it intentionally: the autonomous drifting folks are mad awesome.

@beeoproblem @nonnihil
Dramatically. It swings even further in favour of steering (even at low speed) in those cases. ABS usually won't fire when traction is lost either.

If you need to steer hard in low-traction scenarios (though it's still not recommended), the front wheels also need to be turning in order to achieve that directionality ― so braking hard at the same time makes it even worse.

Of course, that's easier said than done. When you're heading straight for a wall, every instinct is telling you to put the brake pedal through the floor. Takes practice to overcome.

@McCovican @beeoproblem
This (ABS handling of side-loading when approaching the edge of the surface friction envelope) varies a lot by vehicle weight, drive, and model year, so it's worth practicing (esp if you were about to change out your tires anyway).
Also, consider looking for an "advanced driving" course near you (e.g. https://www.driveincontrol.org/ ); they won't generally let you practice with your own car, but you can get a feel for it (and save you an absolute ton on your insurance).
In Control Crash Prevention Training

In Control Crash Prevention - Our Course Consists Of Closed Course, Hands-On Crash Prevention Training, Supplemented By Classroom Discussion. The In Control Family Foundation provides Crash Prevention Training and Community Engagement to educate about the dangers of driving and how to stay safe.

In Control Crash Prevention Training

@McCovican @nonnihil definitely experienced the "ABS doesn't work" in my own driving.

I wasn't expecting steering to be more favorable though. From my own experience steering being favorable is a bit surprising. IME steering without slowing first led to understeer (skidding without turning) on my previous car.

Threshold braking did work but it's definitely not something that comes to mind without practice. Taking your foot off the brake when you want to slow down isn't exactly instinctual.

@nonnihil
Ok, it's wild that this is that different. In Finland, with much shittier on average weather and a pretty hardcore drinking culture (although it's gotten a lot better), but much more aggressive driving training, we have 7.5 fatalities per *billion* VMT.
@SwiftOnSecurity

@nonnihil @SwiftOnSecurity
> 3-4 fatalities per million VMT
Wait, either I don't understand what that means, or that number is wrong. 3-4 fatalities per million vehicle miles travelled would be a 3-4% *annual* chance of death for people driving 10,000 miles per year, which is actually less than the average miles driven. (You can quibble about occupancy, but the median vehicle has a driver and no passengers anyway.)

Does VMT mean something else?

[Edit - ok, 12 fatalities per billion VMT makes a lot more sense - thanks!]

@nonnihil
Wait, I'm seeing 6.9 per billion km on Wikipedia. Are you sure about that number?!
@SwiftOnSecurity
@dymaxion @SwiftOnSecurity
Good catch ; what I get for being so far out of date -- I had a historical number in mind and I multiplied instead of divided to convert to SI.
Per https://crashstats.nhtsa.dot.gov/Api/Public/Publication/813561 looks to me like in the US we're down to 1.2 per 100M VMTs? So that's 12 per billion, so we're in nanofatals.
@nonnihil @SwiftOnSecurity
What does lane divergence mean? People moving between lanes?
@eythian @SwiftOnSecurity
"Lane divergence" is crossing a lane line when you didn't mean to. Modal case, inattentive/sleepy driver drifts left while looking right or something like that.
Lane divergence into a solid object is maybe the single easiest way to get yourself dead on a major road, although into oncoming traffic is more hollywood. Modern safety systems apply a little counterforce to your steering as you cross a line, along with an annoying beep; only the living feel annoyed.