To keep things on the safer side, self-driving cars tend to almost always run on limited speeds, and most autonomous vehicles currently in development are usually tested in public roads.
But what if it's tested on a race course against a real race car driver, all the while maneuvering on hairpin curves at more-than-maximum speeds allowed on public roads?
A team of researchers from Stanford University did just that to find out if self-driving cars can do just as well as humans in unknown and extreme conditions.
Neural Network For Self-Driving Cars
Nathan Spielberg, a mechanical engineering graduate student at Stanford and lead author of the study, and his team created a neural network that is patterned on the same networks found in human brains. This neural network, fed with 200,000 motion samples, can control self-driving cars at high-speeds and low-friction maneuvers.
To try out the real-world driving capability of autonomous cars, the team then brought two test vehicles from Stanford to Thunderhill Raceway in Sacramento Valley. One is a Volkswagen GTI called Niki, and the other is an Audi TTS called Shelley.
First to run the course was Shelley, which is controlled by a physics-based autonomous system that already contains information about the race track, including the entire course and conditions. When compared to the performance of a skilled amateur driver who took the same 10 consecutive trials as Shelley, both showed similar lap times.
Niki, on the other hand, is loaded with the team's neural network system.
"The car [Niki] performed similarly running both the learned and physics-based systems, even though the neural network lacked explicit information about road friction," said a press release about the study.
"In simulated tests, the neural network system outperformed the physics-based system in both high-friction and low-friction scenarios. It did particularly well in scenarios that mixed those two conditions."
A Promising Result But Still Needs More Data
The researchers said that the results of the tests were encouraging, but they stressed that the new neural network still needs further study, as it "does not perform well in conditions outside the ones it has experienced."
Details of the study are published in Science Robotics.
