Dr David Bott, Principal Fellow at WMG

Dr David Bott, Principal Fellow at WMG shares his insights from across the transport sector on how autonomous vehicle safety is perceived, and how we will find the way to widespread acceptance and adoption

In the first article of this series, I discussed whether those who travel understand how safe their chosen form of transport is, and whether that understanding would influence their intention to travel.

In this article, I try to unravel the process by which we can quantify and improve the safety of any autonomous vehicle.

 Who is responsible for safety?

As we consider the advantages of autonomous vehicles, we do have to appreciate the differences between the domains. Although there are striking similarities with the control systems needed, the social context of the ways they have developed leads to differences that need to be considered. With planes, trains, buses and ships, we cede responsibility to a driver or captain, both for arriving at the right destination and the safety of the journey. The vehicle itself is owned by the organisation we are buying the service from. With cars, we are personally responsible for the journey and safety, and we mostly view the car as our own possession.

This latter point makes giving up responsibility for the driving a challenging psychological step. It also leads to an interesting comparison between autonomous cars and human drivers.

We have a clear process to get a person able to drive safely, and we test them on their progress. But who tests the cars when they drive themselves? Is there an equivalent to the practical driving test for a driverless vehicle?

When people learn to be responsible

Let’s start with what we know.

When we learn to drive, there are a few stages. In the first we learn to control the car – to steer, to brake, to change gears and so on. For most, this doesn’t take too much time. In the second we learn the basics of how to drive on the roads with other vehicles. This is achieved by learning and being tested on the Highway Code. This is a basic set of situations we might run into while driving, and the instructions on what to do when we get into them. Having passed the theory test, we go and practice these situations alongside a person with driving experience, dual controls, and the patience of a saint. Statistics suggest this takes about 65 hours in the UK. We then take and (hopefully) pass our practical driving test, indicating we have achieved a minimum safety standard to be allowed to drive independently.

In the third stage, we are approved to drive, but every parent (and passenger) knows that more experience is necessary before a new driver becomes “safe”. If you look at safety statistics for road accidents, although there are other factors, it looks like this can take up to 15 years – by the time a driver is in their 30s, they are about 75% less likely to be in an accident and that doesn’t change much until they reach their 70s. Those early years are taken up with building the experience of driving situations beyond the simple ones contained in the Highway Code and gives a new driver the experience needed to deal with a wide range of situations.

There are clear parallels with qualifying to be a pilot, captain, or train driver. The same stages, more or less, apply – particularly when it comes to the accumulation of experience. Pilots practice in simulators and fly as First Officers (where the Captain also has dual controls!). Future Captains (of ships) work for years on the bridge as Third and Second Officers alongside experienced Captains to watch and learn. It is experience of the wide range of situations that may arise that makes a person in charge of a vehicle optimally “safe”.

There is a problem with cars…

If we compare this process to approving a car (or any vehicle) to drive, fly, navigate or manoeuvre itself, there are obvious parallels.

The basic control part of manoeuvring follows the laws of physics and the principles of engineering, so can be programmed in without too much trouble. If we start with cars, programming the basic 307 rules contained in the Highway Code is probably also not too computationally onerous. It is the experience part of the process where we need to find another way. Having every new car model driving around for many years learning how to drive safely is unworkable. What we need is a means of accelerating this process.

…but computers can help

One potential answer lies in the use of digital twins.

We can build a model of how the vehicle senses and responds to its environment, and then we build a virtual environment for it to operate in. We must populate that environment with things that the vehicle would meet in the real world – for driving it would be other cars, buses and trucks, people, small people (children), animals, weather, and so on. Each domain has its own list!

We then need to cover all the actions they could take that the vehicle would need to react to in a safe manner. The scale and complexity of this synthetic environment means that we only build parts of it because modelling everything would require extremely large amounts of computational power. This means that the vehicle can only be tested in a part of the real world environment, and it is usually the part that it will eventually have to operate in. This is referred to as the operational design domain. Once built, this sort of model can be used over and over again to explore different situations or scenarios within this domain and how the autonomous vehicle would respond.

But how do we know this is how it would act in the real world?

Although we could (and do) use some of the scenarios in the real world, there is an intermediate where the environment is still synthetic, but the vehicle is real. Once again, the parallel with human training is there because these look and feel like simulators!

Using this spectrum of testing, we can explore not just the main designed operating areas of the model but also the cases on the edge of the area. We can take situations which are unclear in the fully computerised model and try them out in the real world. We can take problematic experiences in the real world and play them over and over again, with small changes in the fully synthetic environment to explore what is really going on.

With this data, different domains can share the learnings from near miss or actual incident situations, what was similar, what was not, how could it have been avoided, allowing quicker creation of safer systems.

Check and check again

Used properly, this spectrum of test environments can give us access to a wide range of situations where we need to be sure the vehicle will operate safely. We can also add to the mix by using data from vehicles in the real world, which are increasingly connected, to feed a machine learning system and build even more experience in the synthetic world. We can the take the outputs of that learning and programme it into the vehicle database (or experience bank).

Interaction between the different ways of gaining and testing experience will enable safer operation. The real test will be if this is enough to convince people that driverless vehicles are safe.

Watch to gain more insight into the workshops: https://www.youtube.com/watch?v=1n4BZEJ3Tts&t=10s

Read WMG’s Report “Cross Domain Safety Assurance Framework for Automated Transport Systems”:https://bit.ly/3JcrhclLink opens in a new window

If you are interested in learning more about WMG’s research into autonomous safety and how we work with businesses, please contact wmgbusiness@warwick.ac.uk