Are we imposing unrealistic requirements on automation? Should we be more relaxed in determining an AD system’s capabilities? Or are we simply fulfilling our responsibilities?
Let’s explore the realities a self-driving car would face.
In the world of autonomous driving, the term “safe” is generally defined similarly to how it would be for a human driver: making sure that the vehicle operates in a way that minimizes the risk of harm to passengers, other vehicles on the road, and pedestrians or other objects nearby.
Just like a human driver, a car on autopilot must be able to look at the traffic situation and determine where it can plan a path that avoids potential collisions with objects, seen or unseen, putting the vehicle in a position to repeat the planning successfully. However, a self-driving car must also follow traffic rules, such as speed limits, stop signs, and traffic lights, and be aware of other vehicles and pedestrians around it. That’s how the system becomes predictable and why we’ll gradually feel more confident relying on it.
So far, it seems manageable, doesn’t it? Instructing a car to interpret a sign, stay within lane markings, or adjust its speed according to other vehicles doesn’t appear too complex. Many of us have used driver support systems. Features like adaptive cruise control, automated emergency brakes, automatic parking, drowsiness detection, etc., have been around for years. Some people might even argue that cars are already safe enough.
In fact, why should we pay extra for advanced features we don’t need? If they’re unintuitive or irritating, we’ll just turn them off anyway.
Well, yes, most of us are good drivers. But we’re not flawless. We get tired, distracted, stressed, lost in thought, etc. Automation doesn’t and can, therefore, be the ever-vigilant guardian angel we all need. Using a well-worn expression, a human driver has to be successful every time; an accident only needs to succeed once. What’s more, with increased automation comes the potential to make driving and transportation more inclusive and accessible, regardless of physical abilities or age. This (a tad pretentious perhaps, but still) “democratization” of mobility not only enhances individual freedom but also contributes to a more equitable society.
Moreover, there’s another reality on the horizon: a time when the AD/ADAS discussion becomes less relevant, a time when automation is a perfectly natural complement to human driving. Once that happens, and the car is driving itself for extended periods, every single safety concern will become critical.
For example, when the car exits a highway by itself at high speed, we’ll be grateful for all that relentless product development. And that specific scenario is just a walk in the park compared to a car driving through, say, rush-hour traffic in Paris. That won’t happen tomorrow. Or the day after. But it will happen.
The true challenge for an AD system is managing unexpected situations and unpredictable behavior; automated features must react appropriately to a pedestrian suddenly entering the roadway or a vehicle swerving into its lane. People will always break the rules (voluntarily or involuntarily), and the unexpected will happen. This means we must teach a robot how to deal with erratic human behavior and rare but plausible situations.
In essence, an AV needs the same skillset as a human but without the flaws.
To get to that point, the software must be trained on various driving scenarios, from basic to highly complex traffic situations. This includes not only environmental traffic conditions such as weather, darkness, and road quality but also understanding human driving patterns.
Road data is, therefore, crucial for training self-driving cars. By amassing extensive amounts of real-life traffic information – meticulously curated quality data that we can feed our algorithms – we can analyze any type of traffic scenario to pinpoint how accidents, incidents, and near-misses occur. This information can then be used to develop software that creates safe margins for vehicles, pedestrians, or other moving or stationary objects.
However, anticipating what can go wrong in different ways and verifying that there are mechanisms to prevent it from happening is an extraordinary challenge. The quest for automation that mirrors the capabilities of a human driver is a complex and challenging process. It’s not about creating a fleet of autonomous vehicles in haste, but rather about developing robust, reliable, and safe AD systems through careful, patient engineering.
As we continue to innovate and amass invaluable data on traffic scenarios, we take significant steps towards a future where autonomous vehicles seamlessly integrate into our daily lives. To be sure, it’s a long journey of progress and safety assurances, but still. It’s exciting times.