Instead of experimenting on the roads, a Hungarian company makes the artificial intelligence of the autonomous cars practice with a simulator, while arming them with an outstanding software – reports hvg.hu.
It is impossible to construct the world’s complete road network with every roadblock and accident getting into the database immediately. Yet, Google actually tries to accomplish something like this while experimenting with autonomous cars. Furthermore, not only this market leader American company, but actually all of the companies aim for this. Except for a Hungarian company, the AImotive, and the Israeli Mobileye that was bought up by the American Intel recently. They set out from the fact that a world map requires too many data, it is too expensive and very difficult to keep updated. Because what happens, for example, when the first autonomous car arrives at a temporary lane closure and, on the basis of the map, it thinks that it must not change to the oncoming lane?
Of course maps will also be needed. “The difference is that Google says: I have a three-dimensional model of the whole world, but what I say is: the map is worth just as much as me driving without watching continuously the navigational device” – explains László Kishonti, founder and manager of AImotive. What makes their method different from the one that the cars of Google (and recently of its affiliated company Waymo) use is that, instead of expensive sensors, it relies on only a few cheap cameras and the mechanical brain does not simply practice on roads, but it is learning in a simulator, just like pilots are trained. And it definitely works.
This method would be too expensive for human drivers’ training, but autonomous cars need only one – quite huge – simulator, as the knowledge gained this way can be passed on to all the other cars. All traffic situations can be typed into the simulator. While during those 2 million miles driven by Google only a few interesting things happened to the cars (those probably witnessed a dozen of minor accidents) – said Kishonti adding that testing every situation that may occur would require many billions of driving. And, of course, real emergencies must not be tested in traffic.
The simulator is like a driving game – only they are not humans who are experimenting, but the machine that is made to learn. Just like artificial intelligence usually does, the mechanical driver also learns the way the human brain does – involving neural networks. It does not mean that concepts or rules would be typed into the system, but that, after experiencing many traffic situations, the machine will get to a point when it will “understand” the way of driving.
Cars are dumb right now, they need to be developed further. They do not know many of the situations that may occur, therefore they cannot react properly yet. However, these cars are already able to cope with situations that seem to be difficult for humans. Kishonti gives the example of the six-lane roundabout without a single painted road mark around the Arc de Triomphe in Paris. It is said to be so chaotic that some insurers cast it out of the risks. This kind of herd instinct, the positioning reminding of a shoal of fish can already be handled by the software of AImotive. However, if people were also driving there in the same time, it would become confused; it would be unable to cope with what Kishonti experienced in Paris: scooter drivers tend to kick themselves away from the cars. If the car could sense these people, coming so close, it would be paralized.
“The knowledge of the machine will not be as good as that of the outstanding pilots, but it will run with the reliability of the average driver, and most of the accidents can already be eliminated” – foretells Kishonti. If an accident occurs – and accidents will also happen with autonomous cars – the experience will be typed into the other vehicles. The software of the autonomous cars will be updated regularly the way Windows and other operating systems are provided with safety improvements. According to Kishonti, standards will be provided, which will be demanded by the authorities regulating them.
As data security is problematic in the case of computers (signed by the appearance of the ransomware in the last weeks), it will mean a greater risk in the case of cars. However, problems can also appear without hostile attacks, since computers can freeze. It can also happen that the cameras or important sensors of the autonomous car get dirty, get damaged or do not give signs anymore. Kishonti predicts that, in addition to the big system, a smaller one will also be in service, which will only have one task: to make the car stop in suppliant mode as safely as possible.
Videos and featured image: AImotive