The race towards autonomous cars has cost trillions of dollars and led to substantial developments in AI technology. But as AutoMate’s Harrison Boudakin reports in this article, there is a critical human factor in the development of driverless vehicles that merits a closer look

Modern man is obsessed with designing himself into obsolescence, and in the race to build the first truly driverless car, we have surely found the greatest expression of that fascination. In fact, there can be little doubt: in this second automotive century, the new engineering nirvana is the prospect of eliminating the fleshy, flawed humanoid behind the wheel – and those with the means will stop at nothing to reach this technological peak.
Quite how many trillions have been devoted to this Sisyphean task we cannot be sure, but what we can record is the sheer pace of the advancement in recent years, as automakers and tech-companies alike have staked out their territorial claims deep in the heartland of our automotive future. 
The result is that we now live in the age of the ‘thinking’ car, where the old mechanical art of engineering has been positively spiked with lashings of supercomputing and the crunch of hard data. Complex driver assistance technologies can now be found right at the very bottom of the automotive food chain; meanwhile, apex predators like Mercedes’ S-Class blaze new trails with their increasingly capable ‘autopilot’ systems.
But there can also be no doubt that the road to full autonomy is littered with an innumerable array of obstacles. To give a car complete and unfettered control over its own movement, in every situation, requires still a profound technological leap from where we are right now. And what’s particularly interesting – though not often discussed – is the fact that this leap is going to require rather a lot more human labour than you might have imagined.

As much is apparent when it comes to mapping; specifically, the development of the 3D maps and imagery that help a driverless car locate itself in the world. We may have made enormous advances in sensor technology and artificial intelligence, but the fact remains that a driverless car needs to do more than just react to what it sees; it actually needs to be able to anticipate things too, so creating detailed maps and images of our roads is a task many companies are now focused on. 
At present, numerous mapping companies – such as the Nokia subsidiary, HERE – employ fleets of vehicles equipped with scanning lasers and cameras, which are then driven by people all over the world to capture and record HD imagery. 
But while these companies have logged millions of kilometres already, they face an uphill battle – not only does this effort require an enormous application of human labour and time, it’s also somewhat futile, because roads are incredibly dynamic spaces. You may map a street once, only to find that one week later, the image you generated is no longer accurate. 
In response to this conundrum, one of the leading suppliers of obstacle-detection software, Mobileye, is working on an industry collaboration to begin ‘crowdsourcing’ map data. Mobileye is planning to harvest recorded images from the 15 million vehicles on the road that are already equipped with their semi-autonomous software and hardware.

Mobileye’s parent company, Intel, will then use its expertise in high-speed computing to create a 5G ‘data pipeline’, before fusing the footage with maps already generated by HERE. Eventually, this constantly-updated stream of data will find its way into autonomous cars made by companies such as BMW and FiatChrysler.
Now this all sounds wonderful, but again, there is a key, ‘human’ step in this process that is often overlooked, and yet extremely important.
You see, you can feed an autonomous car with as much data as you want, but it’s essentially useless if the machine isn’t told exactly what to do with that data. Machines don’t know what we know, so it’s up to us to actually “train” them in the art of driving. And that means that right now, companies working on this technology must employ thousands of people to sit in front of computer screens and “annotate” all of the video footage they’ve gathered. 
This process – usually outsourced to China and India – is remarkably archaic by the standards of Silicon Valley, given that the imagery must be analysed frame-by-frame, with human operators adding “notes” to define all the objects that fall within the camera’s scope of vision. 
As a result, every hour of image data generated by a car’s camera can equate to hundreds of hours of human labour, which of course is blindingly expensive. Some estimates suggest that the cost of creating maps for every city in the US alone will run into the tens of billions of dollars. 

Given how limiting this current system is, a number of start-up companies are turning to something called “deep learning”: a new type of machine ‘training’ that essentially allows the computer to teach itself, in much the same way that the human brain does. But this technology is incredibly nascent and has not yet proven itself capable of dealing with the challenging data scenarios thrown up by autonomous cars.
So, until this tech is perfected, we are stuck with a cumbersome, slow and expensive process of manual annotation. No wonder this is a sore point for all the big players in the autonomous game – clearly, the idea that they are relying so completely on a small army of low-paid, offshore data crunchers to build their systems, hardly befits the cutting-edge image they’d like to cultivate.
All this, however, sates an interesting thought.

Arthur C. Clarke once said that any sufficiently advanced technology is indistinguishable from magic, and for modern man, the prospect of the autonomous car has always had this quality. Not only has this kept the public hooked on the idea for decades, but it’s also what continues to make this industry such a source of fascination for investors too.
But as we have seen, the cold, hard reality is that the current generation of driverless vehicles are not actually that magical at all – instead, they remain products of a deeply ‘human’ endeavour. Sure, we may like to think we have advanced computing to the point where the human mind is redundant, but as we can see, this is simply not the case. 
More than anything else then, this proves one thing: that the soft, fleshy, grey bit between our ears is just too advanced for us to be getting rid of any time soon.

AutoMate Training is a world-leading provider of online, video-based technical training, with over 400 modules featuring high-quality graphics, the latest vehicles and real presenters. The AutoMate library covers everything from system fundamentals to advanced diagnostics and technology, including training on Advanced Driver Assist systems. Subscribe today for only $33 per month for professional techs and just $9.90 for trainees.

Visit for a free trial.