Autonomous cars are slowly but surely gaining more appeal among drivers around the world. While a complete self-driving system has yet to be launched, carmakers have been finding ways to make sure it would be plausible and viable in the near future.
Now, Audi is showcasing at the Conference and Workshop on Neural Information Processing Systems (NIPS) in Barcelona, Spain (December 5 to 10, 2016) a new technology that it believes would be one of the keys for autonomous cars – a self-learning system that could intelligently park by itself. Developed by Audi subsidiary Audi Electronics Venture (AEV), this technology is showcased through an Audi Q2 small crossover – more specifically a 1:8 scale model car – that could develop intelligent parking strategies. In a sense, this model car called the “Audi Q2 deep learning concept” is able to learn through trial and error. How does this concept achieve that?
Of course, there is a lot of sensing involved. The Audi Q2 deep learning concept is installed with sensor technology that makes use of two mono cameras – one mounted on the front facing forward and another fitted on the posterior facing rearward. In addition, this sensor technology makes use of 10 ultrasonic sensors installed around the Audi Q2 deep learning concept. These two cameras and ultrasonic sensors transmit data to a central on-board computer that in turn translates these data into control signals for steering and the electric motor.
To demonstrate the ability of the Audi Q2 deep learning concept, Audi featured an area that measures 3 meters x 3 meters, representing a parking area. Without any outside interference, the Audi Q2 deep learning concept autonomously tries to look for a suitable parking space (marked by a metal frame). Here, the Audi Q2 deep learning concept will first – through its sensors – ascertain its current position in relation to the parking space. Then, the model’s central on-board computer will then calculate how the Audi Q2 deep learning concept would drive – in a safe manner – to the selected parking space. As needed, the Audi Q2 deep learning concept would maneuver forward or in reverse, steering accordingly until it has safely parked in the intended space. Amazing, isn’t it?
But what is more amazing is what underpins this ability self-parking ability – something that Audi called as “deep reinforcement learning” or what we call learning through trial and error. Every successful maneuvers and actions of the Audi Q2 deep learning concept is being recorded autonomously by an algorithm logs successful actions, thereby gaining more knowledge through experience. Then, this algorithm refines the parking strategy for the Audi Q2 deep learning concept. Take note that this is a continuous process, which means that the more the system learns, the more it could tackle harder parking maneuvers.
Audi Electronics Venture – based in Gaimersheim, Germany –plans to transfer this technology to a real Audi car. The carmaker, meanwhile, plans to employ this system in 2017 in the central driver assistance controller (zFAS) in the next generation of the Audi A8.