Fraunhofer IESE - Definition Autonome Systeme- Autonom oder hochautomatisiert

Autonomous or merely highly automated – what is actually the difference?

Posted in Autonomes Fahren, Künstliche Intelligenz Published on Authors

Autonomy, automation, autarky, networking… many terms are used in the context of  “autonomous systems”. For example, we already speak about autonomous driving if the vehicle takes over certain driving tasks on its own and without human intervention. Strictly speaking, however, this is only a case of automated, or at best maybe highly automated driving. Because it is so complicated to delineate these terms from each other and because I am often asked for a definition, as Program Manager for “Autonomous Systems” I have compiled delineation criteria in this blog article to provide the view of Fraunhofer IESE on the terminology in this area.

Sie möchten den Artikel lieber auf Deutsch lesen? Hier geht es zum Artikel: »Autonom oder vielleicht doch nur hochautomatisiert, was ist eigentlich der Unterschied?«

What is an autonomous system?

If you search the Internet for a definition of “autonomous system”, the first hits refer to IP networks. The German Wikipedia [1] [2], for example, says: “Autonomous systems are interconnected and thereby form the Internet.” The Fachforum “Autonome Systeme” [Specialist Forum “Autonomous Systems”] [3] differentiates between “digital assistance systems”, “automated systems”, and “autonomous systems”, stating that “a system can only be called autonomous if it can, independently and situationally, achieve a predefined goal without human control or detailed programming.” A matching definition is provided in a report of the expert commission “Forschung und Innovation” [Research and Innovation]: “Autonomous systems can act without direct human intervention, solve complex tasks, make decisions, learn on their own, and react to unforeseen events” [7].

Our program “Autonomous Systems” is also dedicated to systems that act without human direction and still solve complex tasks “autonomously”. But what exactly does it mean to solve a task “autonomously”?

Autonomous through networking

Does “autonomous” mean that the system must accomplish the task on its own, without the help of other systems? Is driverless driving, for instance, autonomous driving only if the vehicle exclusively uses its own sensors and does not have to rely on other sources of information, such as the Cloud, the infrastructure, or other vehicles? Autonomous is often defined as independent and self-sufficient [12], so it seems logical to demand “self-sufficiency (autarky) in terms of information procurement”.

However, many complex tasks cannot be solved satisfactorily in terms of effort, time, and quality without networking. Good decision-making usually requires timely and sufficiently detailed perception of the current situation as well as prediction and assessment of possible future scenarios. Networking plays a crucial role in this: In order to perceive the current situation, it makes sense to use the sensors and the “knowledge” of other systems. In order to take the development of a situation into account already in advance, the future actions of other systems must often be considered as well. This anticipation gets simplified dramatically if the systems network and communicate with each other and even collaborate. An example of such a collaboration is truck platooning, which would be inconceivable without networking. In truck platooning, the individual semitrailers drive so closely behind each other that they appear to be connected by an invisible drawbar. Technical control systems ensure that traffic safety is not impaired.

We therefore consider networking as a generally necessary enabler for autonomy and speak about a closed or an open autonomous system to indicate whether it is networked or not. If open autonomous systems collaborate to accomplish a higher goal, we speak of a digital ecosystem. Robotaxis and public transport, for example, can form an ecosystem in order to incorporate superordinate goals regarding environmental impact and traffic flow. Then it can happen that individual robotaxis do not act optimally with regard to their main objective of getting from A to B as fast as possible because they also support the goals of the ecosystem. If we now add the humans who use the ecosystem, then we speak of a sociotechnical system.

Autonomous and automated refer to our understanding of a system

Both autonomous and fully automated refers to something happening in a goal-oriented manner and without human direction. Nowadays, the terms are often used synonymously – especially in the area of “automated driving”. Various sources [8] even state the fear that the difference will disappear: “I do not believe these differences can be preserved linguistically, even within the profession, the broad misuse and confusion will drown small differences of meaning.” In [4][5][8] and [9], the differences between autonomous and automated are pointed out. If everything has been planned in advance, no matter how complicated it may be, and if we program these planned causal relationships into the system, we speak of automated. But if we do not really capture the causal relationships and use AI approaches to tell the system only indirectly how to behave in a particular situation, then we speak of autonomous.  <…>

Our program “Autonomous Systems“ ties the difference to how the system is understood. If we know for sure how a system will behave in a specific situation, we speak of automation. If we are unsure how it will behave, we speak of autonomy. Autonomy goes hand in hand with the feeling that the system is smart, because we no longer understand directly how it works. If we partially limit this autonomy in order to ensure that no inacceptable risks will emanate from the behavior of the system, then we still speak of an autonomous system.

This view is reflected in the above-mentioned definitions of an “autonomous system”. The Fachforum “Autonome Systeme” [3] excludes “detailed programming”, which would lead to an understanding of the system. In the EFI expert report [7], reacting to “unforeseen events” is demanded and the use of Artificial Intelligence methods is implied.

Not understanding the system means that we partly lose control over the system. This loss of control also fits the meaning of autonomous, because autonomous mostly also means uncontrolled [12].

Autonomous means self-responsible, resp. dependable

However, autonomous also means self-responsible [12]. The point is thus not only that the system does something without direct or indirect human intervention, but also that the human cannot be responsible for the system’s behavior and therefore does not need to monitor it.

This is also reflected in the automation levels for automated driving [15]. The higher the level of automation, the less a human needs to act as a safe fallback level. In the classification of Sheridan and Verplank [14], humans are gradually relieved of decision-making, and thereby also of responsibility, in 10 steps.

The transition from automation to autonomy can therefore also be seen from the perspective of the shift of responsibility from humans to the system. However, a technical system cannot be held morally responsible. It must therefore be built in such a way that it works at least sufficiently dependably at all times. For an automated system, the argumentation of dependability is based mainly on the ability to trace the system behavior in detail, including its defects. For an autonomous system, understanding the system is limited by definition, and defects can no longer be considered deviations from a clearly prescribed (programmed) behavior, but are rather based on decisions made by the system that turn out to be undesired or erroneous.

And this is where we as Fraunhofer IESE come into play: The engineering of dependable autonomous systems is an enormous technical challenge, where information technology is seriously confronted for the first time with systems that cannot be programmed in detail. Evaluating the system behavior in terms of wrong, right, and acceptable goes hand in hand with ethical and legal issues.

As things stand today, the only way to make autonomous systems demonstrably dependable is to limit the “inexplicability”, resp. our lack of understanding of the system, to the part that we do not need for dependability. We do this by building in monitoring systems that check the unpredictable, AI-based behavior and keep it within a safe area. These monitoring systems work in a fully automated, but not autonomous manner, so we understand them and can argue their dependability.

I hope that I have been able to clarify the differences in terminology and that I have given you an understandable definition of autonomous systems. If you have further questions in your company regarding autonomous systems, please do not hesitate to contact me.

References and collection of definitions on Autonomous Systems

Tagged ,