We are “decades away from having completely safe autonomous cars”

The creation of truly autonomous automobiles is “one of the greatest technological challenges that has ever existed”. This is stated by Sebastian Elbaum, associate professor of Computer Science at the University of Virginia, in the United States. Elbaum and the distinguished professor at that institution, Matthew B. Dwyer, have been studying, in recent years, the computer processes necessary to create an autonomous vehicle (AV), safer than one driven by a human being. The result of your work is not at all reassuring.

“We created a model [informático] to try to understand when it is possible to know that we have already done enough tests, when we have already reached a sufficient level of confidence in this type of AV systems to be sure of how they will behave in all types of situations”, begins with explain to DN Matthew B. Dwyer, at the Centro Cultural de Belém, in Lisbon.

The meeting took place during the 46th edition of International Conference on Software Engineering (ICSE24), which took place in the middle of this month, where experts presented their latest work to their peers.

His latest article, titled s³C: Spatial Semantic Scene Coverage for Autonomous Vehicles and signed together with two other researchers, Trey Woodlief and Felipe Toledo, also from the University of Virginia, aims to (roughly speaking) describe how it is possible to process in real time the data captured by a single camera on board in order to quantify, in three dimensions, the viewed environment and, from there, objectively measure the captured information. This for: “Until now, manufacturers use measurements such as kilometers traveled, etc., which in reality don’t say much, because if these were taken, for example, on a motorway, without intersections or pedestrians, the system may even work very well, but once it leaves that environment, we don’t know how it will react,” says Matthew B. Dwyer. “We want to create an abstract mathematical model of the various objects in the environment and their interactions.”

“Until now”, adds Sebastian Elbaum, “there was no way to objectively determine to what extent the vehicle had already ‘seen’ the world in order to make decisions in the future. This one [nosso] This article is a step in that direction – saying ‘wait, you haven’t seen most real life situations, you’ve only seen the most common ones.’ And, of course, it is the less common ones that the car will have difficulty dealing with.”