Developing autonomous vehicles requires planning for the unexpected …
27th Jun 2022 dRISK
…including changes to the Highway Code. Kiran Jesudasan, Transportation Systems Specialist at dRISK, explains why developers need to change their assumptions if AVs are ever to be commercially deployed at scale.
On 29th January 2022, changes were introduced to the Highway Code following public consultation. In essence, a new hierarchy of road users has been introduced, putting those most at risk of a collision at the top. The changes have been implemented after consultation with businesses, organisations and the public.
The notion of putting the most vulnerable at the top of the hierarchy is important for road safety and while there has been some controversy over the new rules, they come at a time when cities in the UK are re-evaluating urban living and how smart city concepts can be applied.
As a researcher in autonomous vehicle safety, I believe there is a fundamental line of enquiry that must be introduced to research. Until now, concepts for safety have been based on what we know to be true at the time, often from lived experience.
The rules of the road have now changed, and it’s vital we understand how human behaviour will change as a result of the law – if at all – and how the development and deployment of self-driving vehicles should be adapted. It’s an important aspect of AV development, and just like the Highway Code has safety at its heart.
Developing safe AVs is complex, and regulators are in a catch 22 state – they need commercial deployments to prove safety, yet they can’t risk safety with a commercial deployment that’s unfit for the roads. What’s needed is a robust set of standards and regulation that accounts for all safety aspects and yet are flexible enough to incorporate the evolving landscape of what is considered truly safe and safe enough. The Highway Code is also a good example of how AVs are only as good as the information they learn from. Up to now, AVs may have been trained on data of road behaviour only before January 29th.
Now they must adapt to new rules and new human behaviours. And in order for the AVs to learn how to behave with these changes their developers need to re train them to do so.
So how can they do this? Let’s step back and look at what has changed.
Firstly, a new hierarchy of road users has been established, where vehicles that can cause the greatest harm in the event of a collision bear the most responsibility to reduce the risk to other road users. This applies to trucks of varying scale, cars, vans and motorcyclists but also considers cyclists and horse riders as road participants who need to minimise risk to pedestrians.
Secondly, pedestrians have been granted increased priority at junctions. All road users need to give way to pedestrians waiting to cross a road or already crossing it. This takes effect on zebra crossings and also for pedestrians and cyclists waiting to cross at a parallel crossing.
Finally, cyclists, horse drawn vehicles or horse riders should not be cut off when turning into or out of a junction, or while changing lane. This applies whether they are using a cycle lane or just riding on the road, and road users should not turn at a junction if doing so causes a cyclist, horse drawn vehicle or horse rider to swerve or stop. Simply put, the changes aim to increase awareness of vulnerable road users and do a lot more to ensure they are safe.
One of the great promises of autonomous vehicles is that they can in theory reduce all human error on the roads. No one company can wholeheartedly say that it is a milestone that they have achieved but over the last few years, through the use of advanced sensors, and in many cases simultaneous 360 vision, the AV industry is exploring exactly how to achieve that elusive goal. In the meantime, however it is the job of regulators and safety authorities to ensure that the AVs on the road do adhere to the rules of the road, including these new changes.
How do they do that? Obviously, it comes down to testing and the new Highway Code changes mean more, and different testing is required. It’s not a change developers will have necessarily accounted for in their plans to launch a commercial vehicle.
Regulators and developers will therefore need a way to look at how changes to laws influence safety assumptions. Take the varieties of junctions in use and the myriad of ways vehicles, pedestrians, cyclists, horse drawn vehicles and horse riders can and do behave at those junctions today.
How will that change in the future? It’s unlikely to be an overnight change, people need to adjust to the law. So, development needs to allow for an understanding that some people will abide by the new rules and others won’t.
One acceptable way to approach the problem is to look at past accidents to inform testing. For example, accurate reconstructions can be built and used to test an AV’s performance in specific and true to life scenarios.
But regulators will still need a way to determine which tests are the most relevant and representative of inherent risks at intersections under the new laws. This is particularly difficult because everyone is adopting the law at different rates. Plus, at one end of the scale there will be notorious junctions that offer a plethora of accidents to simulate, and at the other, more ‘freak’ one in a million chance accidents that an AV could legitimately be confronted with, but no one could imagine.
It presents a challenge – how can developers minimise the number of tests that can be done on an AV stack while covering as many scenarios of what could go wrong? Simulation can be used here to identify areas where the AV is failing considerably quicker than conducting real world tests. However, this line of thinking also helps to introduce a framework to testing too, one where any class of simulated scenarios that an AV fails can be explored more deeply through real world tests.
What can the changes to Highway Code teach us?
Bottom line is that road traffic and design isn’t linear, so it’s easy to imagine that AVs will have to be re-certified against specific tests, determined by a robust data analytics phase that is agile enough to be able to cope with things like Highway Code changes.
It also highlights the risk of bias in the continuous development and learning of AV vehicles. An unprotected right-hand turn with oncoming traffic where several pedestrians are both near and on the opposite pavement, some walking and some just waiting, will be considered high risk. But over training an AV to cope with this scenario to the detriment of a left-hand turn could ultimately yield a less safe AV.
In the same way, AVs need to make fast and correct assessments of risk to a waiting pedestrian when there is a gap in the oncoming traffic. Scenarios need to consider, cyclists as well as motorbikes and vans, different times of day, seasons and weather, and the surprise of a runaway dog.
It’s easy to see then that this one scenario of a left-hand turn can become, in truth, infinite. That is why it is essential to know what type of risky behaviour is inherently present to ensure the types of scenarios used for testing AVs are relevant. The law changes can then be overlaid to these working assumptions.
That’s a critical point. Can someone in a lab set the relevant scenarios? Can they put aside bias of how women use the roads compared to men, or how skin colour needs to be accounted for? And can they build algorithms that can adapt to changes in road laws?
From an AV developer’s perspective, when new highway rules are introduced, the AV stack behaviour needs to be updated to compensate and allow for these new rules. AV developers might look through all the data they have collected and find scenarios that are close or ideally identical to the ones in question, like a cyclist approaching from behind on the left that wants to continue straight while the AV wants to turn left.
They would create several iterations of this to try and see how their AV handles variations. The danger here is that you are limited by what you have seen on deployments. You may well miss fundamentally strange and rare scenarios that are very hard to capture on standard AV deployments.
In a situation such as new rule changes, the chances of missing key scenarios by training in this way are even greater. AV developers will need to expand their scenario pools to include those they have not yet come across but are still grounded in reality.
That’s why it’s critical the AV community works together to establish the practice of leveraging edge cases to ultimately inform safety and architecture of an AV design. Also it will be critical that flexibility be incorporated into the verification process so that when changes like the ones to highway code do happen, they can be handled, and performance verified with elegance. We can no longer only operate on perceived wisdom alone, there are too many accidents to prove AVs have further to go in terms of safety.
Instead, we must find the one in a million chance encounters that have genuinely happened and use them to train AVs to deal with them, creating a bedrock that can be adapted when new rules come in. Until this happens, regulators will be operating with one hand tied behind their backs and AVs will never be deployed on a scale that can transform how we commute and plan urban spaces.