By Carlo Platella
The Indy Autonomous Challenge is the main motorsport competition for autonomous cars, where the experimentation of unmanned driving is pushed to the extreme. Autonomous driving represents an application area for Artificial Intelligence, although not synonymous with it. The technology of convolutional neural networks, in fact, is exploited for very specific tasks, including that of car localization.
The basic modules
The logic for the operation of a self-driving car consists of four layers. At the lowest level is the module used to keep the car along the planned trajectory, which is called Motion Control. In this context, the algorithms fall into the field of traditional control, with PID or MPC logics. Immediately above is the form Planning, with which you set the trajectory to travel through the Motion Control and which makes use of the information returned from the two upper levels. In detail, it is the Target Trackingi.e. the perception of obstacles and surrounding cars, and above all Positioning, the location of the car within the circuit.
Precisely the localization and tracking of obstacles are the modules where Artificial Intelligence finds application in autonomous driving. Explains Professor Sergio Savaresi, head of the PoliMOVE team of the Polytechnic of Milan which takes part in the competitions ofIndy Autonomous Challenge: “With Artificial Intelligence we define convolutional neural networks that, among these four modules, it currently makes sense to use them in the obstacle localization and perception parts. However, it makes no sense to use them in the dynamics part of the vehicle. It's not that it can't be done, but similar results would likely be achieved with much more effort. Even at the Planning level we are not at a complexity that makes its use sensible”.
“We are therefore exploiting Artificial Intelligence, but in two of the four key modules. I don't expect neural networks to be implemented in the Motion Control part in the future, while something could be done in Planning, especially if we run with three or four cars at the same time”. The latter scenario could occur in the newly formed Abu Dhabi Autonomous Racing League, whose inaugural race on April 28th could be held immediately with three or four cars on the track.
The difficulties of localization
Meanwhile, car localization represents the main field of application of Artificial Intelligence in autonomous driving. “Localization is the basic building block without which the machine wouldn't know where it wasae couldn't move. It is a very delicate module,” explains Savaresi. With cars reaching 300 km/h, therefore traveling over 80 meters in a second, miscalculations by a few centimeters or a fraction of a degree risk leading to serious trajectory errors, with disastrous consequences.
The single-seaters of the Indy Autonomous Challenge have a double GPS with RTK (Real Time Kinematics) technology, capable of providing precision down to the centimeter. However, in their absence locating the car becomes much more complex: “When high-precision GPS works perfectly, the problem is relatively simple. The reality, however, is that this scenario does not always occur. In the American oval circuits, the GPS holes they are small and tied to billboards or passing clouds. In Monza however, the entire section of the Serraglio with the underpass presents serious localization problems with GPS. So we had to do without it, relying solely on cameras, LIDAR and radar. This is a big problem, because at those speeds a small localization error is enough to end up in a wall.” This is where convolutional neural networks come to the aid of the teams in processing the information collected by the other sensors, estimating the position of the car with the highest possible precision. The general public's perception of Artificial Intelligence is that of a technology for the future, ignoring however how much it is already being exploited in the present.
#Artificial #Intelligence #role #selfdriving #racing
