Drive with me. If you want to live.

"Right drop me to work then go find the nearest parking spot and wait there. At 4 collect the kids from school then swing by my work to pick me up. Then go get yourself serviced and I will see you tomorrow at 8am."


Speaking to your car like this these days would indicate you are in need of a lot of psychiatric help. However with Google's self-driving or computer driven car and the advent of natural language voice control in all likelihood it will not be long before this level of service comes along. I for one am counting the days.

Here's the thing, humans are a terrible solution to the engineering problem of how to actually control a car. Oh I am sure you are different an attentive perfect driver but well ... it's just you are not. Check pretty much any reasonably large survey of drivers and you find that the vast majority rate themselves as above average drivers. Mathematically they can't be of course. Next discuss driving ability with a few drivers of whom you have first-hand experience of their driving and you will quickly see the extent of the tendency to overestimate personal ability.

The perfect car driver would drive along constantly monitoring all objects likely to come close to them anytime soon. They would guess object trajectories and confirm their predictions multiple times a second updating their predictions if necessary and smoothly altering their own trajectory to minimize any potential risks. Whilst doing this they would be continuously observing all their vehicle's reported signals for warning signs. They would respond to any emergency to the maximum safe mechanical limit inferred from the most current vehicle data available. They would never break the law, get tired or be not totally focused on driving. The inherent biological limitations of being human means of course that no human can tick all these boxes. Despite these limitations humans can be good drivers and on a race track a computer driven car versus a human driven one could be reasonably close but a professional driver would likely win (at the present time of writing). However if this was repeated season after season the human driver is doomed to continuous failure sooner or later.

Human drivers take in and process huge amounts of data via their senses. A decent racing driver uses this to form a model in their mind as to the response of the car. This is often termed as having a good "feel" for the car. The human driver can refine their model but they are stuck with their hardware (human brain v1.0*) whilst very update of the computer can improve both the model and the hardware. As they race both drivers learn from the experience but our human driver is at a massive learning disadvantage as he can't exactly replicate his actions, recall his sensory input perfectly, vary aspects of his driving without affecting other aspects nor make changes that would exceed his biological limitations. Thus the computer can learn more from every previous action and try more things in the search for lower lap times. After sufficient versions (and assuming some governing body doesn't implement blocking rules) the AI's ability will exceed the upper limit imposed by the biology of the driver. There are also fancy computational tricks the computer driven car design team could use like optimization techniques, offline simulation or telemetry crunching that would greatly speed this process up. Of course let's not do us all a disservice here, the human brain is a formidable computer in itself. Even though it is not designed for machine control if the full power of the brain could be suitably interfaced with the vehicle's systems things may well be different but we cannot do this (yet).

Now a game or a race is one thing but surely a daily commute is something else entirely. If you have been involved in computational research for a while you get used to statements like this. That is statements of the form "o computers are great but they will never beat a human at X". The other thing you quickly get used to is seeing such statements proved wrong. A mere 30 years ago most people knew with utter certainty** no computer could ever beat a grandmaster at chess, 20 years ago computers could never suggest what music to listen to based on your personal style, 10 years ago they could never translate on the fly a webpage into something readable (if not perfect), 5 years ago people could never talk to their phone in anything resembling natural language and currently many human drivers "know" computers will never be better commuting drivers than them. Well hell, sorry but just like all the older predictions the drivers will be proved wrong. 

Let me be totally clear, computer driven cars will not be perfect. They will crash. They may even crash in a situation an elite human driver on a good day could have avoided. However we are not trying to make the perfect driver. We are trying to make a driver that is better than say 95% of humans*** and this is much easier. Humans crash a lot, mostly when they make stupid decisions. Statistically 93% of all car crashes are at least partly due to human error and 57% are solely due to driving error. Human drivers make bad decisions far more often than the crash statistics indicate but 999 times out of 1000 it will not matter. You know you should not glance at that beautiful girl/man as you drive down the street. But what will happen if you do. Nothing. You will be fine virtually all the time. Unfortunately just as you get used to this harmless habit the driver in front has an accident and you smash into him and die.

Richard Dawkins might say you died as you formed a false correlation between your actions and a safe outcome. Your biologically evolved tendency to minimize energy expenditure, aka act lazily, on commonly completed tasks got you killed. You will never learn from this event (the dead are very stubborn) and even if you were not killed memories of such a adrenaline loaded event are unlikely to be very useful for learning. A similar crash involving a computer controlled car is somewhat different as once the computer is rebuilt it can use the precise digital memory of the crash to tweak the driving algorithm and lower the chances of it happening again. It is possible that the actual computer involved in the crash might be unfixable but the memory and lessons of the incident can live on even if it has to be retrieved from a mangled black box first. Of course a human might witness or be involved in the crash and drive more safely as a result but slowly he will start to relax into the bad habits. Humans are wonderful at forgetting bad experiences given enough time.

The continual accumulation of experience by computer driven cars is very significant especially for rare events. For example at the start computer driven cars will probably not have detailed experience involving high speed collisions. You can thus make a case that in the beginning when one happens human drivers might be better equipped to take the safest action. Of course it is worth keeping in mind that very few human drivers have any real experience with high speed crashes. Sure stunt/racing drivers have some but driving through a controlled stunt or crashing in a race car on a closed track and being crashing in a road car are utterly different. Heavily whacked cars drive very differently to normal cars and due to the car-accident load of adrenaline pumping through your system so do you. If you think you have significantly increased your chance of survival in a high speed crash scenario because you have played on skid pans or drive at high speed normally you are absolutely dangerously deluded. To be fair to computerkind a computer would never have crashed due to glancing over at a beautiful person. It may monitor the person but they will also keep that safe breaking gap to the next car. Most of the time the gap is pointless and the passengers may moan about it is this that allows a smooth slowdown that saves your life. You probably would not even notice the moment when the choice to let the computer drive saves your life. 

The most common opposition argument to computer controlled cars I hear is from people who say "But I enjoy driving ...". This is in my humble opinion a very selfish opinion. 3000 people a day are killed by driving. Virtually no one chooses to cause a fatal crash they just have an accident or become involved in one. Even the world's best driver can end up involved in an accident. The lifetime odds of dying in a car crash are about 1:240 (UK) or 1:82 (US). Worldwide 33,000 people a day are killed and a further 137,000 injured. Even in the UK which is one of the safer countries on average 5 people are killed per day and over 500 seriously hurt. Every time you get into a car you get into about a tonne of steel which you propel at amazing speeds through a maze of other lumps of steel moving at different speeds. The extreme nature of this experience is totally lost as we do it every day with very rarely even any visible evidence of incidents. When you say "but I enjoy driving" what you are actually saying is "my enjoyment of the feeling of driving is worth the risk of death and injury to myself, my passengers and bystanders". Could you really justify someone who could flick a switch so that when taking their kids to school the risk of their kids dying in an accident was greatly reduced but choose not to as they like the feeling of pushing the pedals and turning the steering wheel?

Oh and if I still have not convinced you robots are safer drivers just consider planes. When planes become too tricky for a human to safely fly machines step in. Autolanding in very bad weather conditions is the classic example. For extreme planes like the Eurofighter or shuttle no human could fly them without at least partial computer help via fly-by-wire. In fact in the case of the Eurofighter it was intentionally designed to be aerodynamically unstable so if the computer and back up systems somehow failed the pilot would have no choice but to abandon the plane. Other potential benefits of computer driven cars only come when you drop almost all of the current driving paradigms. Part two (here) of this pro computer propaganda post will expand on this.




Some passing  points

* Of course the human brain is constantly updated by evolution and has been for a hell of a long time. This however doesn't have any recognizable benefit within the time-scales we are considering here. Nevertheless I am being a tad arbitrary in calling the current version 1.0 but I hope you can overlook this.

** Dr. Hubert Dreyfus is a particularly interesting case he believed and stated "while computers may be able to play checkers, no computer could yet play a decent game of chess". Which was presumably somewhat embarrassing when he himself was publically beaten by a computer shorty after. As for checkers or draughts computers don't actually have to "play" as they have solved this puzzle and thus can't actually be beaten. 

*** You only actually need to be better than the average human to improve things. But as I mentioned people over rate their abilities. Thus machines have to be better at driving than the drivers own perceived ability not their actual ability. Everyone thinks they are above average but most admit they are not in the top 5%.