Basics of Autonomous Driving - Part 1
In this six-part series, central aspects of autonomous driving are examined in more detail - including the degrees of autonomy, decisive innovations in sensor technology and communication infrastructure, the question of social acceptance and finally the ethical implications of the technology.
Part 1: The changing automotive industry.
Everywhere, automobile manufacturers are fighting for supremacy in autonomous vehicle technology. But it is also clear that there is still a long way to go from today's driver assistance systems to highly complex driving platforms that can navigate at busy intersections where machine and man meet. In addition to purely technological challenges, autonomous driving has also triggered a social discourse that is still in full swing.
Autonomous driving is therefore not only about the evolution of vehicles. A major advantage of an automated transport system would be that the number of fatalities could be massively reduced (currently over 25,000 per year on EU roads). In principle, self-propelled vehicles can use a wide range of sensor and measurement technology to detect risks faster and more reliably than humans. In addition, they can communicate with each other to avoid misunderstandings between human drivers leading to accidents.
But first of all, the systems must learn how human road users actually behave. The learning success of modern machine learning technologies such as Deep Neural Networks (DNNs) strongly depends on the amount of data available. Consequently, autonomous vehicle systems must be able to retrieve data from many different sources.
It should be mentioned that machine learning not only benefits autonomous vehicles themselves but can also be used in cloud-based support systems. Vehicles are not only recipients of information, but also transmit an incessant stream of data to these systems in order to further expand the existing knowledge base. The possibility of making the experience gained from one vehicle available to the entire fleet of the manufacturer will result in a swarm intelligence that can be constantly learned and made available via software updates to current and future cars. Tesla has been collecting data on its vehicles since the company was founded, for example, in order to further develop its analysis technologies.
This also gives rise to moral considerations: What regulations should apply to the collection and subsequent use of data so that the privacy of a person is not violated? Who is authorized to collect, analyze, store and share vehicle data? What category of data can it be? Should there be an abstraction layer so that possible trends can be investigated without revealing the identity of individuals? All these questions need to be regulated by law.
Effective use of captured data
A challenge in the design of autonomous vehicles is that a normal car journey is rather uneventful. Accidents are relatively rare given the enormous number of hours vehicles spend on the road. Nevertheless, systems must be able to detect such exceptional situations and react quickly. It would be impractical and, of course, immoral to cause road accidents in order to learn from them. This is where simulation-based training comes in. For example, researchers at Saarland University create complex simulation environments to build systems based on artificial intelligence (AI). These simulations create unusual and problematic traffic scenarios that the software of the autonomous vehicles has to deal with.
Even with extensive databases, it is not always possible to say for simulation-controlled vehicle tests how realistic the respective scenario is. For example, unforeseen sensor failures could induce a real system to assume that no obstacle stands in its way, while simulation with more reliable sensor data shows a completely different reaction. Autonomous vehicles must learn from mistakes - not only from their own mistakes, but also from those of other road users. For this reason, the vehicles will record their experiences and upload them to servers in the cloud, where the AI modules are constantly trained, as long as they are within range of a radio station.
Potentially, a new, minimally safer model for the next day could be installed on vehicles every night. However, the routine recording of daily journeys also has an impact on data protection. For the reasons given above, the content captured may need to be anonymized to prevent the identification of vehicle occupants. This requirement could become even more important if vehicle manufacturers no longer keep their own top-secret development approaches under wraps and disclose them for closer cooperation.
If, on the other hand, everyone continues to act in isolation from each other, each party must develop its own scenarios and hope that those with the greatest relevance have actually been selected. However, if a consensus is reached that general security will improve with more comprehensive data, governments and companies could consider closer cooperation and develop joint training concepts.
How training changes system behavior may depend on more than what sensors communicate to vehicles' AI systems. Overarching decisions come into play, and the foundations on which these decisions are to be made will certainly still give rise to controversy (whose life, for example, should be saved in an acute hazardous situation?). The definition of universal rules is philosophical black ice - but this is a subject we will look at in more detail in a later article in this series.
This article was first published in German by next-mobility.news.