Uncomfortable Autonomous Car Truths
There are many types of challenge in autonomous vehicle development but they can all broadly be summed up by one phrase: Supply and demand.
Yep, when we realised that actually the problems of algorithm development and computer vision are but a tiny blemish on the far deeper-seated problems which blight the sector’s progress, we had to sit down and wonder why nobody else is really talking about it.
Whether talking about people to do the work, technology to make things happen, or circumstance of historical infrastructure approaches influencing how technology is created, the balance of supply versus demand dominates research and development struggles right across the eco-system.
Here are a handful of problems explored…
Autonomous cars love good road markings to understand lanes, but musn’t rely on them because road markings are unreliable on many roads.
So here’s the daft thing, lines are there for humans as a guide to where to drive, to avoid other traffic (lane markings or dividers), provide instructions (junctions, pull-in, etc) or impart other information or influence behaviour.
Autonomous vehicles are largely developed to work alongside human drivers and therefore behave in a human way.
Take away human drivers altogether, have only autonomous vehicles, and you no longer need road markings, but in the short term, autonomous cars are developed to work with or without markings, and with humans – hugely multiplying the problems to overcome. Any one system must still rely on multiple other systems (for absolute positioning) because we must still assume that non-autonomous vehicles will be veering all over the place.
If you want to put your feet up in an autonomous car, will you still be hit by an airbag? Surely the autonomous car should never crash?
The discussion on what people will do with their time in autonomous vehicles has already been going for some time. A mobile office, social hub, media consumption location.. sightseeing base? Anything you can do in comfort on a train or coach should be within reach of a comfortable, specially designed autonomous vehicle. This approach could create its own problems, with a relaxed vehicle designed for lounging, and no desire for restricted movement such as that forced on us by safety belts in cars, what’s to stop us from just relaxing a bit too much and throwing caution to the wind?
Are we about to embark on an era of mixed mode transport, where you can pay extra to drive manually or be chauffeured by your favourite program, or that big brother takes control if you don’t adhere to the local traffic rules and safety straps are automated too?
The safest cars have upwards of 20 airbags, but incorrect passenger placement can put human cargo into harm’s way. Will the safest (autonomous) cars of the future actually do away with the need for additional protection?
Autonomous vehicle design is still trying to herd this particular group of cats. At this point, designers and manufacturer must hypothesise about how the future of vehicle use will be based on new use cases, rather than adapted previous equivalents. It’s a chicken and egg situation entertaining the world’s finest design minds, and leaving a confusing trail of assumptions and dead-ends, all of which may be nonsense.
Satellite navigation (GPS/GNSS) is fairly low resolution, easily spoofed or blocked and fairly unreliable in city or wooded areas, but it’s still the best ‘general location’ navigation system we have
Blending multiple sources of location data has long been a challenge tackled by technologists in industrial automation, and at a few metres per second, is quite straightforward.
Wind up the speed and add interference from the atmosphere, buildings and local plant-life and tuning in to high resolution location mapping suddenly takes a turn into the absurdly complex, requiring high-resolution maps, connectivity, inertia and odometric measurement devices, and time-correction systems to blend with real-time local sensor data from camera and lidar sensors, and it has to work indoors or outdoors, with no help from the infrastructure. Supercomputers are required to achieve the speed of decision making needed – making even entry-level autonomous road vehicles an enormous challenge.
The next generation of satellite navigation, made accurate through a technology called PPP, as well as an industry-wide desire to standardise navigation data and protocols will help, but too much money is spent trying to make a fairly old technology do something it was never designed for – and that is a huge financial burden and technical restrictions on the autonomous vehicle sector, at least in the short term.
Very briefly, satellite signals are also – by comparison to signals generated by your mobile phone – very weak, so it’s easy to replace them. Travel to any major city centre with a GPS receiver and a good computer software interface, and you’ll find spoofing signals from nefarious sources.
All in all, the technology creates a myriad of cost and complexity implications which are an unwelcome distraction.
The most desirable sensors cost too much to get to volume production – someone must make less money, so it’s a race to the bottom that everybody is in and nobody wants to win
Lidar is a low speed, low resolution sensor. In fact, it’s dreadful. It’s expensive. At the moment it’s also mostly mechanical. That’s really only to get past the speed and resolution problems, but in the mean-time introduces the bane of auto-makers – moving parts!
Yes, the sensors that everyone talks about and nobody tells the truth about are actually really bad news. Lidar is, much like satellite navigation, changing – but the supply and demand problem hits here too. It’s basically a camera chip which is collecting encoded infrared beams along with time-of-arrival data, but to achieve close to the sensitivity and resolution required, these devices have been multiplied and then rotated… then multiple units have been added.
We know that the best lidar units (which are still not capable of delivering the data required for full autonomy) cost more than an average car.
One of the best known problems in the autonomous industry, the solution undermines a conventional commercial model – to enable cost-effectiveness, someone has to take a hit.
Newer chip technologies, which will allow non-moving sensors to flourish, are emerging (silicon photomultipliers, which first came to use I the medical sector detecting low energy photons in small numbers).
Innovative new companies are getting funding too, from VC or automaker investment – even some of the older Lidar manufacturers are getting in on the act. Lidar has been around for more than half a century, but it’s been a big of a gravy train for manufacturers pushing products into the space, aerospace, surveying and robotics industries where industry demands were very different.
But to get to the target unit cost which will get auto-makers buying in bulk needs an advance order of millions. Then the market is flooded, and the hay-making days will be over in an instant.
Universities are too slow to pump out the number of researchers and graduates needed to fill industry demand
Most autonomous vehicle companies are recruiting from other sectors and training people themselves – and the digital training ecosystem is getting all the business. Udacity, brain child of self driving car guru Sebastian Thrun, offers various online courses and works closely with manufacturers throughout the ecosystem to ensure training is relevant. A forecast annual intake of 3000 in the nanodegrees first year has swollen to tens of thousands with a new course beginning every month or so.
What happened? It’s easy to blame universities but actually they work on a different timetable from industry, and this particular example is very unusual – the expertise required is very broad, and the level of supervision needed is highly technical and enormously complex (often driven by PhD holders), so the people needed are often very experienced or highly qualified in the contributing areas of expertise.
You might be able to find a postgraduate degree in computer vision, or automotive robotics, but then the capacity of those courses and speed of completion is massively outstripped by currently demand for the technologists, scientists and researchers needed to do the work.
As shared autonomous cars become mainstream, insurance premiums for human driven cars will increase, accelerating the reduction in new car sales
We already know that increased numbers of safety devices in vehicles reduce the net danger that vehicles pose, with some manufacturers already stating mission goals of eliminating deaths caused by their cars, but how does that affect everyone else? How is demand for autonomous vehicles created?
Cultural and economic shifts, such as ‘ecosafe’ driving, increased fuel costs and a generally increased awareness of environmental issues has helped the consumer become more acquainted with fuel economy and in turn put pressure on automakers to make more efficient cars. It’s not unreasonable to assume that the same will happen with safety systems.
These may have been a secondary consideration until more recently, aside from luxury vehicles perhaps where every gadget is desirable and there’s a budget to suit, but as safety systems drop in price and become not just accessible but ubiquitous – it’s within every car-owner’s reach to have safety aids which prevent accidents.
As the numbers are starting to show for newer vehicles from manufacturers such as Tesla, Audi and Volvo, these margins have a net positive effect on insurance premiums. They fall. But the business of insurance is not charity, and that drop needs to be made up somewhere else. Step forward everyone else with no safety systems… up go your premiums!
As vehicles progress through the levels of autonomy, it’ll get harder and harder for them to have accidents, and a greater proportion of L0 vehicles will be visible in crash statistics. Insurance premiums will rise and become a net contributor towards the end of the human-driven car, except for those that can afford it of course, but those are the people that also lead the way!
The real losers will be rural communities, where vehicles are essential, and wages are lower than cities – insurance is a mathematicians’ game, and we are sure you can see why.
Nobody is expecting to make any money from autonomous vehicles for years
On the road at least… there are some exceptions in closed environments such as mines or factories where this tech is much simpler and already mature, but roads are such complex environments and the R&D costs so high that even if you have a full autonomous vehicle, it would still take years of autonomous driving (charged for at ‘taxi rates’) to recover the capital investment needed to buy the vehicle and all the technology, let alone pay for all the research.
Waymo may well be providing pay-for ride hailing services in several cities, but it is a decade away from seeing the return on its multi-Billion dollar investment (just its most recent vehicle orders of Volvos and Jaguars equal that figure alone), and expects to see a faster return not through raid-hailing but through the massively expanded media consumption of ads and content enabled by people being freed up from driving.
In the short term, the folks making money from this sector are the tech suppliers getting rich from all the equipment sales enabled by vast sums of venture capitalist money and R&D investment from much larger firms, consultants talking about the topic, service companies working inside the sector and a smattering of universities and training organisations fighting to keep up – but since consumers are not part of the picture – yet – the next 10 years will see a steadying of that financial picture as investors start to chasten their view on the money and start asking to see some returns.
Only the really big money, Alphabet, Apple and the car makers, along with the players backed by the largest VC funds – have the reserves to wait it out.