Why is it hard to drive - even for artificial intelligence

Google recently launched taxi drones under the Waymo brand, but for now for a narrow category of rides. GM promises to flood San Francisco with autonomous taxis from next year, and Volkswagen plans to begin production of Moia electric robotic vehicles in 2021 to compete with Ford counterparts. But Toyota said that they will leave the steering wheel to a person, but will teach AI to help him drive a car at a completely different level.

It seems that drones for city streets are already becoming the norm. But here's a curious fact - according to the classification of the US Department of Transportation, there are six categories of autonomous driving. At 0, decision-making and control of mechanisms is entirely in humans, at 5 - in the onboard AI. According to this scheme, the best road drone to date, the Cadillac CT6 with Super Cruise, is only category 2. And even then only for the region of 130 thousand square meters. miles for which there are detailed maps. The autopilot in Tesla electric vehicles barely reaches category 2, and Daimler robots show an acceptable level of driving only on autobahns with strict rules.

The challenge of teaching an AI to drive is to evaluate a huge mass of dynamically changing information. For example, it would never occur to a person to think about a tree on the side of the road, wires over the road or the ceiling of a tunnel - we already know that these are static objects, they are not a hindrance. But a dog without a leash near the road is already a threat. AI must notice, identify and analyze each object, without exception, in order to navigate in the current situation.

So what prevents you from putting a more powerful processor and more memory? This is not enough - where will AI get information about what it has never encountered? This person can show intuition and understand that an overly cheerful driver in the next lane can make a dangerous maneuver. And the AI, in this case, must contact that car and request medical data about the person in the cabin. The same goes for the weather, pits and debris on the asphalt, traffic jams, etc. - if the world around is not filled with sensors and does not provide comprehensive information to the drone, it will always balance on the brink of error.