Statistically, aviation is a lot safer than driving. So why not adapt lessons learned from aviation automation for autonomous vehicles and cut the number of road accidents radically?
The company I am founder and CEO of develops software and hardware solutions for autonomous vehicles (AVs). Our work, therefore, exposes the bottlenecks of this evolving technology on a daily basis. In search of inspiration, aviation has been on my radar throughout my journey as it serves as a role model for the automotive industry due to the high level of automation and superior safety standards applied.
Replacing Humans With Algorithms
Commercial aviation is 10,000x less dangerous per mile than car traffic. What's more, 94% of accidents are caused by human error on the ground, whereas that ratio is only 59% in the air. Not only is this because aircrafts communicate with each other electronically via a Traffic Collision Avoidance System, but also because pilots get a rigorous training and they are only allowed to fly if they're perfectly healthy.
Of course it would be naïve to think that providing the same level of training for drivers and bringing their fitness to an equally high level would be feasible, even realistic. That is where AVs kick in, since the only way to radically reduce the impact of the human-factor on accidents is to replace drivers with algorithms. In this effort, we need methods similar to those used in the aviation industry to accelerate the process to reach autonomy.
The Power Of Over-The-Air Updates
One of the reasons flying is relatively risk-free is due to the existence of public databases that store aircraft safety records which stem from onboard black boxes. According to existing protocol, authorities are empowered to ground any airplane where an error is detected, whether it be software, hardware or design. Autonomous driving is headed in the same black box-centered direction. Germany has already announced they are considering requiring AV manufacturers to install such a data recorder.
While this is a great start, there are a reported 10 million driverless cars set to hit the road by 2020. This means there will be an even stronger need for over-the-air (OTA) updates. Just like in the case of smartphones, this will enable autonomous driving software to be regularly updated remotely.
The OTA approach would also prove vital should any fatal technical error occur that could affect a large pool of vehicles. Physical recall of millions of cars would be very costly and slow. Instead of recalls, authorities could order faulty self-driving cars to a permanent halt or to run with degraded functionalities like they do in aviation.
The bottom line is that OTA updates for algorithms will not be for convenience, but a requirement for safety, security, and trust. Needless to say, such updates and online data collection will require proper cloud infrastructure and security, and privacy issues need handling as well.
Simulation Can Make Or Break It
Pilots also tend to make fewer errors than drivers behind the wheel because their training includes sophisticated simulator practices. It's therefore imperative that before AVs hit the road, complex accident scenarios are simulated and algorithms are programmed following the lessons learned from those situations.
When compared to real life road testing, simulations are, without a doubt, more efficient, especially in terms of cost, time and most importantly safety. For instance, with road testing you can't wait for a particular situation (i.e. an accident or weather change) to occur for your modeling purposes. However, with simulation testing variables that are not accessible in real world testing can be further observed, and parameters can be altered quickly, such as changing weather conditions.
But What If An Elephant Blocks Your Way?
Simulated training of algorithms can also come in handy for out-of-the-blue occasions. For example, what happens when an elephant escapes from a circus and blocks your way on the road? Yes, algorithms can see the animal but are not trained to understand its behaviour. Black box data could help reproduce the conditions of the encounter, where simulated training, under a wide range of conditions, can be used to update the algorithms. Since these can be delivered via OTA to all cars within days, public confidence in autonomous technology could be restored quickly.
As you can see, a great number of things must be factored in during the development of driverless technology. I firmly believe that relying on methods, such as simulation widely used in aviation or OTA, will lead to an ultimately safer autonomous driving experience.