Self-driving cars in virtual worlds

Even though a self-driving car in the USA led to a fatal accident and a fall in Nvidia's share price: The company is continuing to research autonomous vehicles at full speed and is sending the vehicle AI to virtual driving school.

Every day, around 3,400 people die on the roads worldwide - one of the main reasons for researchers and programmers to focus on self-driving cars, which are designed to drive through traffic with fewer accidents than human drivers ever could. Nvidia co-founder and CEO Jensen Huang is also certain: "Anything that drives will do it on its own." But before cars, taxis, lorries, delivery vans, buses and tractors can be unleashed on the world without human drivers, manufacturers must first prove that computerised control is safe.

Many test miles required

This requires a lot of traffic data, which is very slow to obtain through practical driving. "Even if a company has 20 self-driving cars in use throughout the year for study purposes, you can only cover one million miles," explains Jensen Huang. "But you need a billion miles to get to 770 accidents."

 Even accident-free special cases such as driving at dusk, when the AI is dazzled by the low sun, are limited to a few minutes a day in the wild. Danny Shapiro, Nvidia's Senior Director of Automotive, says: "At the current rate of learning, we won't see the self-driving car in our lifetime."

AI drivers in virtual worlds

Nvidia's answer to this problem: the car AI learns to drive in a virtual world using a system called Drive Constellation. The easiest way to compare this is to a racing game in which the computer steers the car along a dynamically changing track. This makes it possible to drive for days on end in constant rain or even at sunset, which would take much longer in reality.

In practice, the digital driving school is solved with two servers. The first, called Drive Sim, contains the digital world through which the AI steers the car. Nvidia Drive Pegasus is located in a second server that simulates the artificial driver intelligence. The Drive Xavier successor chip for autonomous driving is intended to combine two Xavier chips and two GPUs from an as yet unspecified Volta successor generation. Both are connected to each other via hardware interloop and update each other 30 times per second.

An initial demo was impressive: the AI uses a wealth of perception routines to analyse what can be seen in front of it and in the exterior and rear-view mirrors and reacts accordingly. At the touch of a button, the sun has disappeared from the horizon and the virtual car is driving at night. "Lidar, radar, we can simulate all of that," explains Jensen Huang. Even physically correct flashing police blue lights, which can then blind the AI driver like a human.

Huang is certain that we will soon see thousands of these virtual driving schools. According to Huang, 10,000 of them would cover three billion miles a year - three times as many as are needed for 770 accidents. Drive Constellation should be ready for launch in the third quarter of 2018.

Source: Heise

Leave a Reply
Related Posts
EN