Stanford's latest autonomous driving tech takes tricky corners like a race car driver
The amount of miles logged by driverless cars are well into the millions, but these closely watched research projects largely ply their trade on sound roads under regular driving conditions. So what happens when things get extreme? Scientists at Stanford University are developing new control software so these vehicles can better handle the unexpected, relying on prior driving experience to remain in control.
"Our work is motivated by safety, and we want autonomous vehicles to work in many scenarios, from normal driving on high-friction asphalt to fast, low-friction driving in ice and snow," says Nathan Spielberg, a graduate student in mechanical engineering at Stanford and lead author of a new paper describing the new software. "We want our algorithms to be as good as the best skilled drivers – and, hopefully, better."
Rather than program autonomous cars to navigate their environments by relying on sensor data, Spielberg and his colleagues instead built a system that leans on physics and recent driving maneuvers instead. This meant using 20,000 physics-based trajectories to build out a model and use that as a foundation for a neural network that incorporates data from recent driving experiences.
Except these weren't regular driving experiences – well, not for most people anyway. The team had two of Stanford's autonomous vehicles, called Niki and Shelley, a Volkswagen GTI and Audi TTS, respectively, complete laps at an ice-covered test track near the Arctic Circle and the Thunderhill Raceway in California.
These exercises were designed to explore the limits of friction, which has a direct relationship with how much a car should brake, accelerate and steer when performing emergency maneuvers. With the data then incorporated into the neural network, the team says they wound up with a promising new control method for autonomous vehicles.
"With the techniques available today, you often have to choose between data-driven methods and approaches grounded in fundamental physics," says J. Christian Gerdes, professor of mechanical engineering and senior author of the paper. "We think the path forward is to blend these approaches in order to harness their individual strengths. Physics can provide insight into structuring and validating neural network models that, in turn, can leverage massive amounts of data."
The system was put to the test again at Thunderhill Raceway, where Shelley and Niki's autonomous capabilities were pitted against an experienced race car driver, who they were found to perform "about as well" as. The researchers are buoyed by the results, but say the system does have some limitations in that it doesn't perform as well in conditions it is yet to experience. But as autonomous cars continue to gather data, they are hopeful the capabilities can be expanded.
"With so many self-driving cars on the roads and in development, there is an abundance of data being generated from all kinds of driving scenarios," Spielberg says. "We wanted to build a neural network because there should be some way to make use of that data. If we can develop vehicles that have seen thousands of times more interactions than we have, we can hopefully make them safer."
The research was published in the journal Science Robotics.
Source: Stanford University