The art of learning

Forward, reverse, forward, reverse again. The miniature white Audi Q2 manoeuvres back and forth in a three-by-three-metre walled-in area. After a few moves, it stops perfectly in the parking spot it was aiming for. Doesn’t sound very unusual? Well, look closer and you’ll realise what’s so spectacular about this feat. The little Audi Q2 has performed its task entirely without human intervention. No remote in sight. This performance was staged as part of the Audi showcase at NIPS (Neural Information Processing Systems) in Barcelona, one of the world’s major artificial intelligence conferences. 

Don’t let its compact size fool you; inside, this miniature car is nothing short of remarkable. It’s the Audi Q2 deep learn-ing concept, and Audi is using it to demonstrate its vision of piloted parking. What’s so amazing about this presentation? Throughout, new challenges are repeatedly set for the miniature Audi. Before each round, the car and its target – the parking spot – are repositioned within the demarcated area. That means the little car has to adapt to a unique situation each time and develop intelligent parking strategies. 

The underlying technology, deep reinforcement learning, is a type of machine learning based on a sophisticated neural network – similar to the human brain. And, like humans, the Audi Q2 deep learning concept also acquires knowledge by trial and error. An algorithm autonomously recognises successful actions and uses them to continuously refine the parking strategies. Both successful and failed attempts are assigned a score. Based on that, the algorithm in the Audi Q2 deep learning concept keeps fine-tuning its parking strategies over the course of several million iterations, i.e. repeated attempts. 

Machine learning is an essential stepping stone on the road to piloted driving. That’s because, for a car to accurately assess complex situations in confusing, real-life urban traffic and then initiate and execute the right manoeuvres, it must be able to think for itself, monitor actions and learn. One of the Audi engineers’ next steps will be to transfer the parking space search process to a real car. After all, as the saying goes, practice makes perfect.