Over the weekend, I got a bit ranty on the Twitterz about autonomous vehicles, taking issue with a variety of things, including some of the numbers used to promote “autonomous vehicles save lives” claims, and the infallibility of autonomous vehicles because of, you know, “AI”.
Noting a link from the Overspill blog today from the New York Times – America Is Now an Outlier on Driving Deaths – showing how the US is one of the laggards in the “Deaths per billion vehicle miles traveled” stakes, I thought I’d grab together some of the links I shared previously just in case I ever get round to starting to dig into this a bit more.
As weak support for my belief that “convenient” numbers are used to make “justifiable” claims about the significantly better safety record of autonomous vehicles, I’d shared the following links without digging in to the them (they’re starting points…):
- on fatality and accident rates differing in different countries and on different road types:
- WHO /Global Health Observatory (GHO) data on Road Traffic Deaths
- UK Department for Transport National Statistics: Transport Statistics Great Britain: 2016, for example TSGB0801 (Reported road accidents and casualties, population, vehicle population, index of vehicle mileage, by road user type), TSGB0802 (Reported road accident casualties by road user type and severity), TSGB0803 (Reported accidents and accident rates by road class and severity), TSGB0809 (International comparisons of road deaths: number and rates for different road users, by selected countries), TSGB0811 (Motor vehicle offences: findings of guilt at all courts fixed penalty notices and written warnings, by type of offence)
- on possible side effects, weak signals on some of the things the lobbiests might go after to simple the problems faced by autonomous vehicles, perhaps to help provide better accident stats in early days of any roll-out to provide further “evidence” for why autonomous vehicles are better:
- a Wired article Maybe It’s Time to Cede US Freeways to Driverless Cars on a white paper from a VC firm proposing an ““autonomous vehicle corridor” replacing the I-5 freeway between Seattle and Vancouver”; I wonder if access to things like bus lanes or AV lanes on motorways is also being lobbied for in the UK?
- a proposal from the US National Highway Traffic Safety Administration on Federal Motor Vehicle Safety Standards; V2V Communications. “V2V” in this case being ‘vehicle to vehicle’, and relating to such things as the transmission of Basic Safety Messages (BSMs) “which contain data elements such as speed, heading, trajectory, and other information”. These are presumably easy to build in to autonomous vehicles and are likely to also benefit other new vehicles with sensors already in place for lane tracking, cruise control etc. I’m saying nothing about the privacy implications that spring up around the local broadcast of this data… It also sets up the model for arguments where technologies used to support autonomous vehicle operation become mandated as required on human riven machines. (Presumably, because machines don’t need seatbelts, US drivers will continue to assume they don’t need to wear seatbelts either…)
I’m not saying AVs don’t / won’t have better safety records than human driven cars under particular circumstances, I’m just wary about the basis of any comparisons made and how the numbers are chosen to make the AV case more attractive…
I haven’t had time to dig into the numbers, but it’d be interesting to see how different breakdowns of the actual numbers compare to headline figures (selected how?) proffered by the AV lobbiests. I suspect they go for a convenient average, where Anscombe’s Quartet is more the reality…
One thing I suspect the lobbiests will claim is improvements in safety projected from the implementation of a range of technologies, not all of which are necessarily limited to autonomous vehicles, such as BSMs that could be used by human driven cars in support of things like lane following and advanced cruise control. I also suspect AV lobbiests will compare road deaths per billion km in US now with projected road deaths from AV vehicles with BSM implemented everywhere in some point in the future. Which would be an unfair comparison (assuming that human driver rates could also be improved with the adoption of some of the same yet to be adopted wider vehicle safety technologies).
I also haven’t seen any simulations of road accidents on motorways and dual carriageways for different road and traffic conditions and different mixes of “human” and “autonomous” drivers?
On the matter of “AI, therefore good”, I am circumspect:
- the trained models used to generate control signals for AVs will almost definitely include biases we haven’t identified yet; I also wonder about the range of different models required for different national driving styles and road types (English country lanes are not the same as wide US boulevards). I also wonder about how the models will need to evolve over time, as folk change their own behaviour with respect to AVs and as more AVs hit the road and more “machine assist” technologies are built in human driven cars. And what happens when you go from one regime (high density AVs/machine assist) to another (high density human drivers);
- tech gets hacked, for profit, fun or out of sheer maliciousness. For example, hacking the environment to fool machines, as in recent demonstrations of Slight Street Sign Modifications Can Completely Fool Machine Learning Algorithms; and folk are already working on proof of concept attacks based on recreated models (for example, Generating Adversarial Malware Examples for Black-Box Attacks Based on GAN. Will particular paint jobs be made illegal becuase they’re confusing to computers? I’ve also been reading The War Magician: The True Story of Jasper Maskelyne, which makes me thing it’d be interesting to see what a magician turned hacker could come up with…!
- people are also playful and will likely try to game or taunt AVs, as in James Bridle’s Autonomous Trap 001 ;-)
I also wonder what protocols will exist when meeting an AV on a single lane country road, for example, or situations of four cars meeting at a mini-roundabout (an occurrence oft-encountered, and that can lead to frequent delays, on the Isle of Wight…!). Who will give precedence to whom, and how will it be signalled? If autonomous vehicles are only allowed on particular categories of road, then does that makes it more likely that we will also end up with roads/lanes where only AVs are allowed (to “make it fair”?!).
I imagine there are lots of other concerns – but also lobbiest claims – out there for the reading of in the evidence of the recent Lords inquiry into Connected and Autonomous Vehicles.