A new system could improve the way autonomous vehicles cope with demanding and unfamiliar conditions. The neural network enables self-driving vehicles to combine physics data with previous driving experiences to drive safely even in conditions the vehicle has not seen before.
The neural network has been developed by Stanford University. Like human drivers, autonomous vehicles regularly face demanding conditions, such as roads that are slippery due to ice or mud. In this case, human drivers can refer back to driving experiences they have done before to adjust their driving behaviour and drive safely in these conditions. Autonomous vehicles, however, do not have this ability, as these vehicles typically base their decisions on pre-programmed rules and observations using sensors and cameras.
Combining physics with data-driven methods
"With today's technologies, you usually have to choose between data-driven methods and methods based on fundamental physics," explains J. Christian Gerdes, a professor of mechanical engineering at Stanford University and involved in the development of the neural network. "We think the future lies in combining these methods so that the individual strengths of both can be exploited. Physics can help structure and validate neural network models so that they can then be deployed to process massive amounts of data."
The neural network is deployed to determine what resistance a road presents. This is an important piece of information that determines, among other things, how fast a vehicle can safely brake, accelerate and take corners. "Experienced drivers, especially if they have racing experience at amateur level or higher, are able to operate a vehicle safely at the limit. In racing, this is demonstrated by setting low and consistent lap times. In a critical situation, however, this means being able to deploy all the resistance offered by the road to avoid an accident," the researchers write in a scientific paper in Science Robotics.
In analysing road resistance, the neural network deploys information gathered during previous driving experiences gained at Thunderhill Raceway Park racetrack in Willows, California, and at a winter test facility. This experience was gained with two self-driving cars from the university: an autonomous Volkswagen GTI named 'Niki' and an autonomous Audi TTS named Shelley.
Safety
"Our work is motivated by safety. We want autonomous vehicles to operate in a variety of situations, ranging from driving under normal conditions on asphalt with lots of friction to driving fast on ice and snow with little friction," said Nathan Spielberg, a postgraduate student in mechanical engineering at Stanford University. "We want our algorithms to perform as well as the most skilled drivers - and hopefully even better."
The neural network was compared to an experienced human driver during a test. In the first phase of the test, Shelley drove with a physics-based model that was pre-loaded with information about the track and conditions. Using this system, the autonomous Audi TTS managed to set lap times comparable to those of an experienced amateur-level driver for 10 laps. Next, the university's Volkswagen GTI 'Niki' was sent onto the track with the new neural network. During this test, this neural network was not given any information about the resistance of the road and had to determine this resistance itself based on previous driving experiences. Despite this, Shelley and Niki recorded similar lap times during the test.
Simulated tests were also carried out. During these tests, the neural network outperformed the physics-based system in both high and low road resistance scenarios. Especially in situations where both conditions are combined - such as a cold winter day when on some parts of the road are passable and other parts are slippery - the system performs well, according to the researchers.
Network further development
The researchers stress that their neural network only performs well in conditions it has previously experienced for itself; with situations where there is no previous driving experience, the network cannot cope. However, the researchers point to the huge amount of data collected by autonomous vehicles today. If this data is used to further train the neural network, it is possible to further develop the network to cope with a wide range of conditions.
Author: Wouter Hoeffnagel
Source: Stanford University
Photo by: Kurt Hickman
