A little different from other self-parking research vehicles we've seen, Audi's Q2 deep learning concept is a 1:8-scale model that relies on machine learning to help it hunt down its mini parking space and maneuver itself inside. Cameras and ultrasonic sensors route data through a central computer, which controls navigation, and the car learns from trial and error, continuously refining its self-parking capabilities. This type of learning will be integral in giving real autonomous vehicles the ability to handle complex driving situations safely and efficiently.
Audi is showcasing its scale self-parking Q2 at the Conference and Workshop on Neural Information Processing Systems (NIPS) in Barcelona this week. The car demonstrates Audi's ongoing research in artificial intelligence and machine learning.
NEW ATLAS NEEDS YOUR SUPPORT
Upgrade to a Plus subscription today, and read the site without ads.
It's just US$19 a year.UPGRADE NOW
A pre-development project of Audi subsidiary Audi Electronics Venture, the Q2 deep learning concept uses a sensor set of front and rear mono cameras and 10 ultrasonic sensors to navigate around the 3 x 3-meter model parking area, identify the parking space and maneuver itself inside. The sensors route data through the central computer system, which then controls the steering and electric drive motor.
"The model car's parking ability is made possible by deep reinforcement learning," Audi explains. "In other words, the system essentially learns through trial and error. To begin, the car selects its direction of travel at random. An algorithm autonomously identifies the successful actions, thus continually refining the parking strategy. So in the end the system is able to solve even difficult problems autonomously."
The car's parking ability improves over time through this process and errors decrease. Audi identifies this type of artificial intelligence as essential for developing autonomous cars that can effectively analyze and react to complex situations, such as urban traffic.
"In one year a 16-year-old driver might encounter 1,000 situations with a four-way stop," Peter Steiner, managing director of Audi Electronics Venture, was quoted in a Nvidia blog post earlier this year (Nvidia is one of Audi's technology partners). "But our systems will learn from hundreds of thousands, even millions, of such situations that can be stored, analyzed and improved from, so these cars can learn even better than a human being can."
Audi's next step will be to transfer this self-parking tech to a real research car. The next-generation A8 will feature deep learning-based software when it debuts next year.
More on the technology can be seen in the video below.