How to create an autopilot for heavy vehicles using artificial intelligence?

Question

Automation

Hello.
Today, the headlines of various foreign publications are full of the creation of cars that can drive with little or no driver participation. Large corporations spend a lot of money on these purposes. I wonder if it is possible to achieve similar results using small blood. I came across an interesting article www.bloomberg.com/features/2015-george-hotz-self-d... and realized that, apparently, it is possible.
It would be interesting to communicate with people who have an idea about this, it would be interesting to create a prototype.
It looks curious to build an autopilot based on artificial intelligence, which will self-learn.

Answer 1

Vitaly Pukhov is absolutely right. The solution of the guy from the article is smart, effective, but not universal at all. Most importantly, it targets American drivers, American roads (on which Hotz trains autopilot), and the mild climate of San Francisco, where it snows once every ten years. In addition, he uses rather expensive hardware for training - the construction of a similar prototype will cost you ~ 300,000 rubles, not counting the cost of the car itself.
Just so you understand: Google strives for autonomy in > 99.99% of all possible driving situations, Tesla - in about 98-99%, Hotz's technology is fundamentally limited to a deliberately lower percentage, since it does not have:
1) the advanced mapping system that Google and Tesla use (read the resolution they remember the road surface at) - Hotz will never be able to implement this on his own;
2) group training (fleet learning), when each machine gains the experience of all machines at once - this is also not feasible without proper infrastructure;
3) an advanced system for recognizing objects on the road, because in “2000 lines of code” this is not feasible in principle. It takes years to develop such systems, no matter how genius you are, because if the system correctly determines the object in 99% of cases, then in every hundredth of a real situation there is a risk of hitting or colliding with another car, and we do not need this.
Moreover, Russian roads, drivers and weather conditions combine just those last fractions of a percent that are most difficult to work out. We still have federal highways that people cannot drive on . Where is the work!
Most likely, Hotz has so far implemented the following functions:
1) route calculation and tracking based on GPS and Google maps;
2) following the road markings (keep in the lane);
3) following the car in front and keeping the distance in the lane;
4) recognition of road signs and, possibly, traffic signals;
5) an emergency stop in front of obstacles and a controlled stop in case of a critical decrease in controllability;
6) some aspects of learning (presumably the system remembers and averages Hotz's response to changes in certain road conditions - the appearance of an obstacle, contextual road signs, etc.).
While we know that the Hotz system behaves confidently during the day on a familiar road in good weather, we do not know at all how it behaves on an unfamiliar road and in conditions of limited visibility. Hotz himself acknowledged in the article that the autopilot only worked the morning before the interview.
What KamAZ is doing there is still childish talk, and AI doesn’t smell at all there. By October, they only implemented driving along a pre-programmed route at a limited speed, copying the leader's trajectory (which is essentially the same), remote manual control and, presumably, a stop when an obstacle is detected in greenhouse conditions (it is not yet known if this function automatically works Or is it also the result of remote control). That is, the mostelementary tasks, which, in a good way, are implemented by the efforts of a couple of smart programmers in a week or two at most. You can’t even go out to the city street with them, but you can’t even dream of the M56 and the like. This suggests that they have already run into one of the difficult-to-implement tasks and have not yet advanced enough in solving it to talk about it.

Answer 2

russianaicup.ru - virtual machines with AI, but only the authors of these machines self-learn.. :)

Answer 3

Forget about little blood, especially in Russia. The development of such a system takes thousands of hours, dozens, and even hundreds of high-level specialists. You can write a crooked trough from which anyone will suffer sooner or later, or you will have to spend a lot of money on development, testing and resolving issues with the law.

Answer 4

To begin with - to write a system for recognizing road objects in real time - to build a virtual model based on its data. In a virtual model, lay a route, calculate the probability of collisions - rebuild the route in critical situations.

Answer 5

Who is interested in participating in such a project - write to
[email protected]