Yandex taught self-driving cars to drive like humans using trained AI

«Yandex connected a transforming neural network to the autopilot system, instructing it to plot the trajectory of the car. The model was trained in tests with high-class drivers – it will help autonomous vehicles react more flexibly to the road situation, and the driving style will become more “humane”.

Image source: Yandex

To create an array of training data, Yandex attracted high-class drivers who completed emergency driving courses and confirmed their skills. They moved around the city, and the system recorded information received from sensors about the road situation and the driver’s actions. Thanks to this, the neural network has learned to respond to known and unknown situations that may arise hypothetically. The new planner has already begun to be used in practice: when traveling around the city and in the simulator.

Test results showed that a trajectory planner based on neural networks turned out to be more effective than a system based on algorithms. In cases where the old-generation autopilot could not avoid a collision, the new one often found a trajectory along which the car avoided the obstacle without getting into an accident. Yandex has been working on autopilot technologies since 2017 – such cars are being tested in Moscow, Innopolis and Sirius. And in October of this year, the company’s unmanned truck delivered goods for Yandex Market from Moscow to Tula along the M-4 Don highway for the first time.