Researchers Dániel Fényes, Balázs Németh and lead researcher Péter Gáspár from the ELKH SZTAKI Systems and Control Laboratory recently published their paper “Novel Data-Driven Modeling and Control Design Method for Autonomous Vehicles” in the journal Energies. In the paper, the researchers present a new model developed for the design of the control system of autonomous vehicles.
One of the biggest challenges in the development of level 5 autonomous vehicles – where the vehicle can drive completely independently – is to create a control system capable of taking the high number of different variables involved into account. After all, driving a car is not a linear system in which you can travel at a given speed – in real traffic, the speed, direction and environment of a car are constantly changing, which means you need a control system that can adapt to all of these different variables in real time.
When developing this type of control system, the experts involved often begin with the operation of advanced driver-assistance systems (ADASs). On one hand, this is a logical approach, as ADASs have been on the market for a long time, but on the other hand, the reliability of an autonomous system can only be 100 percent if no one has to intervene in the process at any time, even in an emergency. Based on data from millions of experiments, SZTAKI researchers have successfully created a model that can help develop autonomous systems.
The researchers used big data-based analysis methods to build the model. After analyzing the data collected from the experiments run, it was decided not to build a neural network-based model, as it might become too complex to design a guaranteed high-quality, safety-critical management system during learning. Instead, the researchers worked with model structures with variable parameters and control models based on these parameters. A control model created in this way can simultaneously adapt to the effects of the constantly changing environments as the vehicle travels, while providing a guarantee of high quality characteristics at the same time.
The completed algorithm was tested by researchers using the CarSim car simulation software. The testing included running the program on a virtual model of the Michigan Waterford Hills racetrack. Compared to the nominal procedures available so far, the version based on the new modeling procedure resulted in a more manoeuvrable autonomous vehicle model. The differences were particularly noticeable in the case of extreme maneuvers and sudden turns.
A huge amount of data is required for this procedure. And even with the data, researchers had to work hard to fit the algorithm to a given task. At the same time, the results show that the new method can create a more efficient and powerful self-management model than nominal procedures.
The English publication is freely available at this link.