The artificial intelligence that will allow the developers of autonomous cars to try them inside a computer

The autonomous cars manufactured today are able to navigate and drive on the freeways, but no autonomous car can turn a crowded Times Square. An autonomous car that drives well in Palo Alto will not do well in Bangalore. To achieve this, we developed a simulation system based on artificial intelligence (AI), says Coganta Chairman Alon Atzmon at the Israeli Chip Club

The main scientist

Alon Atzmon, the chairman of Coganta, presented at the meeting of the Israeli chip club Coganta's simulation system designed for manufacturers of autonomous cars to allow them to drive billions of hours inside a computer, in order to prevent accidents in reality: "Coganta is building a platform that makes the quality control system for autonomous cars, something that does not yet exist . To enable an autonomous car to be put on the road, the manufacturer needs to accumulate 11 billion kilometers. So far, for a whole decade of autonomous car development, all companies developing driverless cars have collectively reached 5 billion kilometers."

Recently, there has been a wave of accidents caused by a crazy failure of all the AI ​​systems - both the cameras and the optical radar systems - Lider. Atzmon fears that the two fatal accidents that happened in March - to Uber and Tesla cars, will set the entire industry back years. In the event The best - one or two years. This is a loss of billions of dollars for the car companies.

The autonomous cars manufactured today are able to navigate and travel on the freeways, but no autonomous car can turn a crowded Times Square. An autonomous car that drives well in Palo Alto will not do well in Bangalore. To achieve this, we developed a simulation system based on artificial intelligence (AI).

"Our system contains four layers or modules. The first module is a static module - a model produced with the help of maps. The second layer is a dynamic layer that simulates the behavior of other drivers on the road. This layer is based on real information - for example, we placed cars in a simulation of Bangalore and discovered edge cases Many you need to know how to respond to."

"The third layer is the simulation of sensors, and here we use deep neural networks, because it is impossible to take the imaging devices and do tests of the interaction of the car with a drop of rain. It will take a whole day to produce a frame, and we need a system that works one to one - an hour of simulation simulates an hour of driving The fourth layer is the cloud, where the code for use by the car companies is located, when Amazon's servers can be connected to facilitate the run."

In response to questions from the audience, Atzmon presented some such extreme cases - "For example, if it rains on the road at night, because then a reflection is created and you see 2. If you are driving on a road that has been driven in front of you in snow, how can you distinguish between the tracks of the wheels and the lanes? We are able to overcome all these distortions - and to create a real picture, as well as to describe exactly how it will look like traveling in Munich and how - in Bangalore. We are able to produce accurate models of cities A centimeter from many sources including Google maps, OPEN STREETNET and more. And if someone wants to travel tens of millions of hours in one night, they can even connect 5 million Amazon computers for the purpose of running."

According to Atzmon, Coganta employs ten deep learning experts. "These are rare skills, since the field is relatively new, but the best have come to us, one brings the other because things happen here."