In the realization of Autonomous Driving, Waymo relies on ChauffeurNet, an Artificial Intelligence intended to imitate humans.
The robot taxi service of Waymo should learn to drive better, what a special Artificial intelligence has programmed: ChauffeurNet. The model should be the human being.
The vehicles of Waymo One are not very popular in Arizona, recently there were even attacks on the vehicles. These attacks result from dull hatred. But many travelers in the US state of Arizona are excited about the restrained driving style of the cars.
That may be one of the reasons why Waymo is now working hard to improve the driving style and has created ChauffeurNet. Another reason was that one at real test drives complex driving situations, which presented the car with insurmountable problems.
Up to now, artificial intelligence has been trained with 30 million examples of training material. But that was not enough to prepare the vehicles for specific situations. It was able though stop signs or lights to interpret correctly, however, deviations led to renewed difficulties.
Therefore, one has taken the human driving behavior as a model. Because people can handle unfamiliar situations while the robot car tends to stop. So you have a recurrent (recurrent) neural network (RNN) worked out that called ChauffeurNet.
The ChauffeurNet is divided into two components: FeatureNet and AgentRNN. The FeatureNet is a neural network that examines the input data and extracts important information such as the road, traffic lights, speed limit, current and past position of other vehicles. The AgentRNN uses the data to calculate the further procedure. The concept is reminiscent of that RSS process from Mobileye,
The ChauffeurNet also uses the data of the training, which was extended by added disturbances. This is called imitation losses. This forces the model to deviate from what has been learned. It tries to solve the problems in which it avoids collisions and at the road marking oriented.
This approach worked on the test site, but there is still a long way to go in the vehicles. So it’s up to the system to teach why seasoned flesh and blood trainers behave the way they do. So it’s about the ‘why’ and not about as many examples as possible.
The current system for autonomous driving in the Waymo vehicles is already an Artificial Intelligence, but operates rule-based. With ChauffeurNet the existing AI should be extended.
source (English)
About David Flora
I’ve been writing about the topic of Autonomous & Connected Driving since 2011 and also talk about it on other sites like the Smart Mobility Hub. I studied social sciences at the HU Berlin and since 2012 I am a freelance journalist. Contact: mail@autonomes-fahren.de