Summary of Key Points
Recently, many car owners have observed, and in some real accidents, a phenomenon that goes against common sense has emerged: many vehicles advertised to have advanced driver assistance systems are capable of accurately avoiding pedestrians crossing the road or even children suddenly appearing out of nowhere. However, these systems seem to ignore suddenly appearing animals such as sheep or cows on the road, as well as unmarked roadblocks or concrete piles placed temporarily, and end up crashing into them. This revelation sheds light on a widespread misunderstanding about intelligent driving technology. The ability of such systems is not about being able to recognize and avoid everything; rather, it reflects a series of practical trade-offs made by car manufacturers in terms of training data, allocation of computing power, risk assessment rules, and legal considerations. The primary goal of these systems is to ensure the highest level of safety for pedestrians, not to become “omnipotent drivers” capable of handling every unusual situation. The intention behind such technology is never to show off.
---
Detailed Explanation
Why Avoiding Pedestrians is a “Basic Requirement,” While Avoiding Animals is a “Less Common Task”?
You can think of the AI in intelligent driving systems as a driver learning to drive. The “test questions” for this AI are the training materials provided by the car manufacturers. In traffic laws around the world, protecting pedestrians is a top priority, so car manufacturers provide the AI with millions of images of pedestrians in various scenarios—during the day, at night, in the rain, with umbrellas, pushing strollers, using crutches, or even children suddenly appearing from between cars. Even cosplayers in strange costumes or delivery workers wearing all-black raincoats are included in the training data, allowing the AI to recognize them even with its eyes closed.
But what about sheep? Sheep are rarely seen on city roads, and most car owners never encounter animals during their daily commutes. The AI’s training data may contain less than 1% images of sheep, with many of these images being screenshots from online videos. Some images even mistakenly identify standing dogs or white burlap bags as sheep, so the AI is not well-trained to recognize these unfamiliar objects and naturally fails to respond correctly when faced with them. The same applies to temporary roadblocks: random piles of concrete, sand-filled bags, or partially damaged traffic cones have no standard appearance, and such “non-standard” objects are not included in the AI’s training database.
Intelligent Driving Systems Have a Hidden “Priority List for Safety”
Many people assume that the computing power of these systems is unlimited and can scan everything on the road at once. However, this is not the case. The computing power of current automotive-grade AI chips is limited. They need to monitor multiple vehicles, lane lines, traffic lights, and road signs simultaneously. If the system were to treat every small object on the road as a potential danger and react with sudden braking or sharp steering, the car would drive erratically, which could cause passengers to feel nauseous and increase the risk of rear-end collisions.
Therefore, car manufacturers have established a clear priority for the AI’s actions: pedestrians have the highest priority, followed by normally moving cars, electric vehicles, and bicycles, and only then by animals and stationary obstacles on the side of the road. Many advanced driver assistance systems simply assume that immobile objects on the side of the road (such as road signs or utility poles) are fixed structures and rely on the driver to avoid them. This is to prevent the system from making random steering movements that could lead to accidents involving other vehicles.
Previously, there have been numerous reports of intelligent driving systems crashing into parked fire trucks or malfunctioning vehicles due to these priority rules.
Humans and AI Have Different “Cognitive Logics”
Car owners often wonder why the AI can’t recognize a large, visible sheep. The reason is that humans rely on common sense to make judgments. For example, if I am driving in a rural area and see a white shadow by the road, I can infer it might be a sheep and slow down in advance. However, AI lacks this common sense; it only compares the current object with the pixels in its database. In dark conditions, a standing sheep and a white plastic bag or a bag of fertilizer may look very similar, and the AI cannot distinguish between a living creature and an inanimate object. Some sheep may also be covered in mud, making them indistinguishable from shadows on the road. Humans can use context and experience to fill in missing information, but AI does not have this ability, and unknown objects essentially “do not exist” to it.
Car Manufacturers Avoid Creating “Omnipotent” Systems for Fear of Liability
Some people accuse car manufacturers of lacking technology, claiming they can’t recognize sheep or roadblocks. In reality, adding tens of thousands of additional images of sheep and roadblocks to the AI’s database would not be a technical challenge and would only cost a few hundred thousand dollars. The real reason car manufacturers refrain from doing so is fear of liability. According to current laws, the driver is fully responsible for all accidents, even if the accident is caused by a system malfunction. If a system were set to react excessively to small objects (such as a cat or a chicken), and it caused an accident with a nearby pedestrian or electric vehicle, the manufacturer would face significant legal and reputational risks.
Therefore, car manufacturers focus on ensuring that the system can handle the most common and dangerous scenarios (such as avoiding pedestrians) with absolute reliability. For less likely scenarios (like hitting sheep or non-standard roadblocks), they intentionally limit the system’s intervention, leaving the driver responsible for those situations. This is a matter of risk distribution, not an indication of technical limitations.
In summary, modern driver assistance systems are essentially advanced “passengers” that help you avoid pedestrians by applying the brakes. Their purpose is not to impress with their ability to handle all types of situations but to reduce the risk of accidents in the 99% of daily driving scenarios involving pedestrians and other vehicles. For the remaining 1% of rare scenarios, it is still up to the driver to pay attention to the road.