Major automobile manufacturers are eagerly moving forward with autonomous driving — which is nothing short of a colossal project that will profoundly change our relationship to the car and its use. At present, autonomous driving is transitioning from Level 2 to Level 3. The fundamental difference is that at Level 3, the vehicle completely takes over driving the car during suitable conditions, and the driver no longer has to monitor the system continuously. Since the car can independently perform all driving tasks with a high degree of reliability and can react fully automatically and appropriately to the surrounding traffic situation, the driver is able to turn away from the immediate task of driving and the surroundings. Several automobile manufacturers and suppliers are developing a specific, and ambitious, function that automatically pilots the vehicle up to a maximum speed of 80 miles per hour.
Level 4 systems are also under development; these are innovative concepts that enable completely driverless mobility on public roads during more challenging conditions, such as moving at low speeds that, for example, are experienced in cities. Tests are underway in urban areas.
To be considered a fully autonomous, Level 5 vehicle, the automobile must be able to self-drive without restriction on all roads, in all traffic situations, in all lighting and weather conditions, and over the entire speed range — at least as well, but probably considerably better, than a human driver. Where the threshold of social acceptance lies and when such vehicles will be available for purchase are still unknown factors.
Autonomous cars require enormous amounts of data
Notwithstanding the legal and ethical problems, the biggest challenge in exceeding Level 3 is to prove that the system can behave safely and appropriately in all conceivable situations within the specified boundary conditions, called the operational design domain (ODD). To do so, auto manufacturers will have to collect, process, store, retrieve and manage enormous amounts of data throughout the different phases of function, system and software development. Case in point: 100TB to 200TB of data are generated alone every day by a test vehicle outfitted with reference sensors.
To prove the quality and reliability of an experienced human driver in an autonomous, Level 5 vehicle, an estimated 10 billion miles of test drives will be required. But it is only possible to test-drive a few million miles on real roads at a reasonable cost. If done with vehicles equipped with a reference sensor system for measuring distance — such as a camera, radar, or LIDAR (Light Detection and Ranging) laser scanner — about one hundred million miles of useful scenarios can be generated using open-loop simulation (resimulation). In this method, simple parameters of the real traffic situation are varied.
The problem, however, is that even with a hundred million miles, we would only have run in just under 1 percent of the data required to secure a Level 5 system, and rare and complex traffic situations would hardly be covered.
Covering the necessary number of traffic scenarios can only be accomplished with the help of a closed-loop simulation. In contrast to the open-loop simulation mentioned earlier, in which data is collected on real roads with a sensor rack, everything takes place in the virtual world. First, the vehicle, the sensors, the driving behaviour and the entire environment must be virtualised well enough so that the remaining 99 percent of the test drives, including the rare, complex and dangerous traffic situations, can be simulated virtually.
This is the current challenge that needs to be overcome. With an autonomous driving platform, the huge amounts of data can be efficiently processed, and the driving algorithms trained and tested. In addition, all data is consistently available to the new autonomous driving development processes. Such a platform is already being used for series development by several automobile manufacturers. In the current expansion stage, the focus is on highly efficient management and analysis of the data volumes generated, as well as training, which is made possible only by this data. Recently, two important services have been added to this platform: semiautomatic video labelling, and the simulation and virtualisation program. These solutions alone bring us several steps closer to the next level of autonomous driving — and more are sure to follow.