Bits and bytes instead of metal and gasoline: control units with artificial intelligence are a key technology in self-driving cars. Digital systems for the mobility of the future are already being trained in the present day. They learn like little children.
Too good to be true? A construction site suddenly appears in the road, but the person in the driver’s seat does not have to touch the steering wheel. The car steers clear of this hazard with the help of artificial intelligence (AI). The AI initiates the steering movement.
Although the robotic car is still driving around the construction site exclusively on a test track, automobile manufacturers such as Daimler, BMW, and Audi are already demonstrating with their research vehicles what will be possible on the roads of the future. The topic of self-driving cars has picked up speed. Automobile manufacturers have been improving assistance systems for years. The arrival of artificial intelligence has accelerated this process. In the current transitional phase, it is already possible to perform many driver desires before the driver speaks – instinctively, so to say. The AI notes the mood of the driver using data on how he/she is steering or stepping on the gas/brake pedals. The onboard system then offers to play classical music to help a stressed driver relax, or something peppy to reinvigorate a fatigued driver.
Modern vehicles have become data centers on wheels. They have hundreds of chips, sensors and mini-computers, which control all of the vehicle’s electronics. The automobile is being taught to read the mind of the driver. “We want to simplify mobility with the help of artificial intelligence by following a coherent, proactive approach to shaping the mobility of the future”, says Demetrio Aiello, Head of the Artificial Intelligence and Robotics Labs at Continental. In addition to manufacturers, automotive suppliers are developing more and more competencies in this area; they are also convinced of this technology’s importance. Continental has a team doing research on AI in three locations across the globe and is conducting joint research with Oxford University, among other things.
Demetrio Aiello distinguishes between intelligent and non-intelligent tasks. Today, many tasks requiring intelligence burden drivers; with the help of artificial intelligence, we can relieve them of some of these tasks. One of the intelligent tasks is driver modeling: an AI system observes and creates a digital model of the driver, the purpose being to react to his or her condition when necessary and thereby make mobility safer.
Electronics are still far inferior to humans: they can only do what humans have programmed them to do. To make robotic cars ready for production, machines will also have to learn. This happens when their algorithms are given great amounts of data. Only when the AI has a wide range of data available can it actually drive around construction sites or anticipate the sudden appearance of a child playing behind a parked vehicle and initiate an evasive maneuver. This is particularly difficult in traffic because there are always new situations.
Instead of programming a machine explicitly for individual situations, it is given the tools for understanding its surroundings. These tools are neural networks replicated by a computer. They learn like children, who explore and learn about the world through their sensory organs. Image data in which objects are first classified by hand is then fed into the system. Once the computer has seen enough images of pedestrians, it can automatically recognize humans in new images. The more training it gets, the better the algorithm.
The Daimler Cityscape research project, in which videos of typical street scenes in 50 different cities were automatically evaluated by such a system, shows how well the training works. The computer had to differentiate between 30 different object classes (pedestrians, cyclists, cars, houses, traffic lights, traffic signs). Pedestrians were identified even if they were largely hidden by parked cars.
A challenge for manufacturers is to reduce the amount of time needed before new technologies are available for use in vehicles. Up to seven years may pass between designing a new vehicle model and its production start. The development cycles of new control units last around two and a half to three years, according to Aiello. “The availability of sufficient processing capacity and the possibility of over-the-air updates can reduce development times”, says Aiello.
According to the Continental manager, it is only a matter of time before machines will be at the same level as an excellent driver. The AI will then drive cars without having decisions affected by emotions, stress, alcohol, or drugs. “This will enable entirely new freedom in automobile design, because the role of the car will change from being the safest and most comfortable ‘driving machine’ to being a mobile living space.”, says Dr. Steven Peters of Technology Management Digitization and Corporate Research at Daimler AG, referring to the “F 015 Luxury in Motion” research vehicle.