Une time again, the automobile was at the party at the Consumer Electronics Show (CES), the high mass of tech which has just closed its doors in Las Vegas with, once again, artificial intelligence (AI) at the heart of innovation.
The cockpits of the future offered by the manufacturers created the “buzz”. These cockpits, for which comfort and safety are priorities, will tomorrow incorporate facial recognition for unlocking the doors, smart headlights to improve visibility in all conditions, sensors for the automatic detection of potentially forgotten children the rear (900 deaths from heatstroke in the United States since 1990) or monitoring by artificial vision and physiological sensors of the driver’s state of wakefulness, the direction of his gaze, even his emotional state, to avoid accidents as much as possible and facilitate the work of the driver assistance systems.
But whoever talks about AI and the automobile thinks of the autonomous car. The first on-road demonstrations date back to the 1980s, with the work of pioneers such as Ernst Dickmanns and Dean Pomerleau. This technology is now arriving in production vehicles. Europe distinguishes 5 levels of autonomy: at levels 1 and 2, the on-board software brakes, accelerates or keeps the car in its lane, but it is content to assist the driver who remains solely responsible. True autonomy begins at level 3, where part of the responsibility is entrusted to the software (and therefore indirectly to the manufacturer), to overtake a car on the highway for example. The driver must nevertheless be ready to regain control in the event of an alert launched by the vehicle. At levels 4 and 5, no more driver, the car is fully autonomous. Only Honda and Mercedes-Benz have been selling Level 3 cars for less than two years, while Tesla is still at Level 2. Time will tell if – and when – cars offering Level 4 or Level 5 autonomy will be deployed.
The whole panoply of AI
Why does autonomy remain so difficult at a time when ChatGPT can, it is said, do the homework of high school students? The autonomous car uses the full panoply of AI by integrating learning, perception, navigation and planning modules to solve problems as difficult as pedestrian detection or the real-time fusion of data acquired by cameras. , radars and lidars capable of mapping the road in three dimensions. Integrating them into a system that is reliable in all circumstances is even more difficult: the supervised learning behind the latest advances assumes vehicle deployment conditions similar to those encountered during the training phase of the software driving it. However, here, the sources of variability whose effect is multiplicative range from the urban landscape and the weather to the traffic and the “agents”, human or artificial, which evolve around the car. Predicting the behavior of these, even over a few seconds, in the presence of such uncertainty, remains a challenge for modern AI.
You have 16.67% of this article left to read. The following is for subscribers only.