A broken-down truck on the shoulder of a highway. A person behind the truck walks out into the lane. Suddenly, a car passes at high speed.
In this scenario, a fatal collision is inevitable. It all happens too fast. No automated system in the world can react fast enough. Therefore, the only way to avoid a tragedy is to not end up here in the first place.
This means slowing down when passing, possibly even changing lanes. This, in turn, requires acting several seconds before passing the truck. For a non-distracted human, this wouldn’t perhaps be asking too much. We have a remarkable ability to adapt to various situations and handle unforeseen circumstances.
But what about a robot?
Can autonomous driving systems learn to achieve such a high level of predictive ability and prioritize safety? Can they be genuinely precautionary in approaching traffic? For such systems to be… well… human-like, they must learn to prepare for the unexpected.
We believe they can. And as an AI company exploring every new emerging angle of deep learning theory and practice, we believe they can learn that from us.
One of the primary challenges in designing AD systems is teaching them to learn and adhere to human values and preferences. One such principle might be the boy scout motto, “Be prepared.” Or, if you’re more pessimistically inclined, “Always prepare for the worst.” But it’s a balancing act; if your machine is instructed to be overly cautious, you’ll never get out of the driveway.
In any case, regardless of how you phrase a robot’s dynamic, predictive, adaptive, precautionary, preparatory, anticipatory, fail-safe, or… human-like… approach to potential danger, the point is this: autonomous driving systems must demonstrate their ability to handle a wide range of scenarios – even those that were not part of their training data. In a nutshell, this is what machine learning is all about – teaching a machine to generalize and make decisions beyond the information it was fed.
We, too, must find ways to generalize the car’s actions – helping it uncover every pattern hidden in the data – so that it can eventually handle complex situations it might not have encountered before. To come as close as possible to ensuring the safety of an autonomous driving system, we must therefore consider the following basic tenets:
- AD systems cannot only rely on traffic rules and regulations for guidance.
- AD systems must be resilient to human errors and assume that not all road users comply with the same rules.
- AD systems must adhere to the same principle we are taught in our first driving class: don’t drive beyond your capabilities.
These fundamental principles are necessary for constructing a safe AD-driving policy. In short, cars must be able to adjust their driving to different circumstances. While it is still too early to say whether robots will outperform human drivers in real-life traffic, automation clearly has the potential to reduce the number of traffic-related deaths and injuries significantly.
But what is it, then, that supposedly makes automation so superior to human driving? And what kind of equipment is needed?
Let’s take a quick look at one case in particular: darkness.
In the US, approximately 6,500 pedestrians are killed by vehicles yearly, with 80% of these deaths occurring at night. With the right combination of hardware, including cameras, radar, and lidar, as well as software, this number can be dramatically reduced since cars with the right equipment can see far better in the dark than humans.
Good performance in darkness is crucial for semi or fully-automated systems. Lidar, for instance, effectively detects objects and structures without requiring additional light sources, making lidar-based safety a crucial enabler for AD rather than simply a crutch to lean on.
By also integrating a high-definition map, a car can predict upcoming road conditions and accurately identify its location, navigating properly even without lane markings. Autonomous vehicles rely on maps to perceive and make sense of their environment beyond the range of their sensors and cameras. This approach – this combination of sensors, cameras, and HD maps – creates the best conditions for safety and convenience for autonomous driving.
Moreover, not only do these technologies outperform us when it comes to absorbing the environment – at a great distance, in the dark, around corners, over the hill, and so on. We can also depend on an AD system not to get stressed, depressed, angry, distracted, or tired. It doesn’t drink and drive. It doesn’t speed or break the rules to impress its robot buddies.
And even if a human on red alert can also avoid imminent danger on the road, the system is always on, always watching, always vigilant.
Realizing the vision of human-like automation (with super-human powers of perception and route planning) requires vast amounts of quality data, cutting-edge AI-powered software, the right hardware and other helpful tools. However, technology alone may not be sufficient to eliminate car accidents worldwide; driver education, public awareness campaigns, and infrastructure improvements are also essential in enhancing safety on our roads.