Autonomous systems of the future. Classification, features and requirements

Autonomous systems are known today largely due to the latest trends from the automotive industry. In fact, automated systems of varying degrees of autonomy are an integral part of future developments and development plans for many areas of activity. The article presented by the authors Werner Damm and Ralf Kalman from the magazine Informatik-Spektrum Edition 5/2017 presents various industry norms and standards, as well as describes the functional capabilities and requirements for methods, processes and tools for the development of appropriate software.

What is the benefit of autonomy?

How much should a technical system and how autonomous it can be?
Today it seems that there are no boundaries for the implementation of more and more advanced autonomous systems. We are on the verge of introducing technologies to the market that independently build complex relationships of the surrounding world based on the data provided, automatic identification of objects and information from sensors at various levels. All this is used to get an accurate digital representation of reality for the realization of the task. Systems are introduced that are capable of analyzing the possible further development of events in the surrounding reality, which goes far beyond human analytical abilities. Systems are being implemented that independently plan and implement tasks without requiring any support from outside. These systems are endowed with human cognitive abilities,

The recently published German high-tech strategy report from the German government shows many possibilities for autonomous systems. Among them are all types of "Smart Systems", such as Smart Mobility, Smart Health, Smart Production, Smart Energy, whose intelligence is realized on the basis of the above possibilities. They are capable of real-time creating a digital picture of the world, processing data from a variety of information sources, and organizing the joint operation of millions of subsystems in such a way as to ensure the successful fulfillment of goals, such as, for example, optimizing the use of resources. The benefits of this can be applied in many areas of public life: health and transport, energy consumption, productivity and quality of products, prevention of natural disasters and collisions of various vehicles. Philips, for example, when using special wearable sensors for postoperative observation of patients, expects a reduction in postoperative cardiac arrest by 86%, and through “smart” monitoring of critical health parameters in outpatient treatment, a reduction in its cost by 34%.

Automated control systems have been around for quite a few years. Automation allows you to effectively use technology without the need for manual intervention. Typical tasks of automated control and equipment setup are represented as control circuits, for which mathematical models are created and implemented in the form of electronic devices and software.

The modern development of cyber-physical systems goes far beyond these limits. Combining IT with embedded control systems and dynamic interaction with each other ensures their work together through heterogeneous data interfaces. As with automation in the 1980s, autonomous production promises increased efficiency, productivity, and quality.

Such developments are carried out in many areas of application of technical systems. Although their application scenarios are different, in the field of software, common problems can be identified and generalized methods for solving them are described. Examples of such methods will be presented in the last part of this article. Of particular interest is the use of self-learning systems. With them, the potential possibilities of autonomy seem limitless, because it is possible to recognize the initially unknown, affecting the operation of the system, the artifacts of the surrounding world, and study the dynamic models relating to them. Thus, new, previously unpredictable possibilities of using equipment are being opened.

The potential market value of technologies arising from these developments is estimated at hundreds of billions of USD. In particular, the study of the European project Platforms4CPS provides the following data:

  • By 2035, autonomous cars will account for 10% of all sales. This corresponds to approximately 12 million units and a market volume of 39 billion USD.
  • The air traffic control market will grow, according to an estimate, from 50.01 billion USD in 2016 to 97.3 billion USD in 2022. At the same time, the average annual growth rate will be 11.73%.
  • The global market for robotic aviation expects an average annual growth of 17.7% over the next decade, so that in 2025 its value will reach 7.9 billion USD (according to Markets and Markets ).
  • The volume of the drones market is estimated at 13.22 billion USD and by 2022 should reach 28.27 billion with an annual growth of 13.51%.
  • The market of unmanned vehicles is estimated from 437.57 million USD in 2016 and up to 861.37 billion in 2021 with an annual growth of 14.51%.
  • The market volume of autonomous underwater vehicles will increase from 2.29 billion USD in 2015 to 4.00 billion in 2020, with an estimated annual growth of 11.90%.
  • The Industrial Internet of Things (IIoT) expects growth from 110 billion USD in 2020 to 123 billion in 2021.
  • The technology market for wearable gadgets has a volume of 28.7 billion USD. Gartner predicts that this market will grow annually by an average of 17.9% between 2015 and 2017. At the same time, the segment of wrist mobile gadgets with an annual increase of 30% is the most growing.
  • The market for microgrids was estimated at 16.58 billion USD in 2015. In 2022, Markets and Markets expects its growth to 38.99 billion USD with an annual growth of 12.45%.

Thanks to the development of technology, new types of products and services with a high level of automation are emerging on the modern market. This raises the question of in which areas such developments really make sense, and what impact they have on society.
In the conditions of constantly increasing level of autonomy, the quality of interaction between man and technology will definitely change. Today, a person acts not only as an end user, but, in many cases, as part of a management system ( human-in-the-Loop). Autonomy creates a trend that establishes the interaction of man and technology at a higher level of abstraction. The autonomous system gives a person the opportunity to familiarize himself with a part of his digital vision of the world with the help of suitable abstractions, such as, for example, virtual reality technologies that are relevant for solving a specific problem at a given time. Conversely, a person can easily influence complex processes within the system through intuitive human-machine interfaces. This communication, accompanied by a rising level of abstraction, requires, in turn, a certain level of qualification and training. At the same time, jobs for low-skilled staff will disappear as unnecessary.

Continuous use of a large number of data sources will greatly increase the risk of their exposure. The architecture of networked distributed systems will place extremely high demands on its protection in order to avoid the catastrophic impact of possible cyber attacks aimed at disabling individual constituent components.

With growing autonomy, the question also arises as to what values ​​the underlying decision-making process should have, and whether they correspond to our own. Based on this, the European Parliament in its resolution of February 16, 2017 decided:

  • use the principle of transparency, which implies that it should always be possible to establish the principles and arguments behind every decision made with the help of artificial intelligence that can have a significant impact on human life;
  • it should always be possible to present the computational algorithms of a system using artificial intelligence in a form understandable to man;
  • Progressive robots must be equipped with a so-called “black box” that records data about each transaction made by the machine, including logic that has contributed to making a decision.

Finally, due to the upcoming entry into the market of autonomous unmanned vehicles, it is necessary to revise the laws on liability for arising offenses.

These topics thus go beyond their purely professional sphere. How should autonomous systems be designed so that they bring not only economic benefits, but would also be positively perceived by society? These problems should be the subject of study in computer science. It is time to rethink the existing processes and design techniques in which the social impact assessment of autonomous systems being developed should be included on an ongoing basis.

Autonomy classes in various industries

The most famous example is autonomous vehicles in the automotive industry. Many manufacturers announced the release on the market of the corresponding cars in the next 3-4 years. However, the support systems that are already available today make it possible to realize amazing things. Despite this, the path from partially automated driving (some manufacturers also speak of “manned” driving in this case) to fully autonomous driving is still very far away. With partially automated (corresponds to the 2nd level of automation for SAE) the main responsibility lies with the person, and he must be able to independently intervene in the process as soon as possible. In addition, the possibility of using such systems is limited to a strictly normalized environment (for example, driving on a highway). At the highly automated driving level (automation level 3 for SAE), the driver is allowed to pay attention to other things, that is, the software guarantees complete driving safety or, in the event of any error, puts the system into a safe state, for example, stopping the vehicle on roadside. Fully automated cars (automation level 4 SAE), which do their job completely without the driver’s help, represent the highest degree of autonomy,

A significant influence on the development of this industry is, first of all, not the desire of ordinary people to transfer control of their car to other hands, but the needs of new transport companies in relevant services, opening up new market segments or offering more efficient and fast public transport within populated areas. In freight transportation, automation allows you to unload the driver, who can devote the released time to other tasks and will thus work more productively.

In the railway and, in particular, in the underground transport, some processes are already automated. There is a simplified model here, since the system works on a homogeneous landscape, where there is no intersection of transport routes and many of the routes are isolated from each other. On the other hand, to this is added the superior system of management and coordination of processes, which is why the International Union of Public Transport ( UITP)) included in its classification of the superior monitoring and control system. The automated train system contains the following three components: safety, train management and train monitoring. Safety is controlled by keeping the distance between the trains, as well as controlling their speed. The control ensures the movement of the train according to the schedule and regulates, for example, the opening and closing of the doors of the cars. Supervision of trains controls, in turn, all routes and the entire infrastructure and transmits the relevant information to the control center.

Such a system can be most easily implemented in the metro based on the homogeneity of vehicles and the isolation of the infrastructure. However, the relevant concepts can be transferred to other areas of railway transport, up to large sorting stations. At the same time, there are still problems in observing and controlling the movement of international transport or because of the complexity of the environment, such as, for example, the movement of suburban trains at railway stations of various types. The engine advancing the automation of railway transport is the high economic benefit of the proposed solutions, achieved, for example, by saving energy with concerted processes of acceleration and braking in one transport network.

In air transport, automated flight control has been used for a long time. For the UAVs, used mainly for military purposes, the level of autonomy was increased in terms of self-planning tasks and mission management. The ten autonomy levels of ALFUS (Autonomy Levels for Unmanned Systems) use three projections to characterize the capabilities of the system: independence from human intervention, complexity of tasks, and complexity of the environment. Together they characterize the ability of battery life. When searching for technological solutions for a higher degree of autonomy, such topics as behavior in a group, adaptive communication between devices, and self-study are also added, which, as yet, have not touched on the other taxonomies mentioned above.

In production, automated processes are standard with the introduction of programmable logic controllers (PLCs) in the 1980s. Such processes, however, have little flexibility and are focused on mass production. Individualized production or market-driven changes in the product portfolio lead to costly retrofitting of production lines and equipment re-equipment. In the process of developing digital technology and based on the concept of Industry 4.0individualized production seeks to achieve the same level of efficiency and quality as in mass production. At the same time, it should automatically adapt to changing conditions and new production goals. The Frauenhofer Research Society offers 5 evolutionary stages that accompany this development. First of all, it is required to ensure the collection and processing of production data. This will be the basis for supporting systems that assist in work and in decision making. At the third stage, the integration of production stages into a single data exchange network and their integration with each other provide the necessary conditions for optimizing the entire system as a whole. To increase the elasticity of production in the fourth stage, the system requires the ability to transform and reconfigure. And at the last fifth level, the production system must be able to organize itself. To date, production systems have settled at levels from the first (production data collection) to the third (production united by a network of common data, such as in the production of cars). To move on to the next stage, as a rule, a complete restructuring of the entire production architecture is required, which, respectively, is costly.

The levels of autonomy of all the listed applications are shown again in the table, and an attempt is made to present similar degrees of autonomy from different domains at the same level.

Autonomy levelMotor transportRailway transportAviationProduction
0No automation"Rides as he sees"Data collection and processing
oneAuxiliary systemsAuxiliary systems
2Partial AutomationAutomated security systems with a driverLimited controlWork in a single network and integration
3Conditional AutomationAutomated safety and operation systems with driverReal-time status diagnosticsDecentralization, adaptation and transformation
fourHigh automationUnmanned operationError, breakdown and flight conditions adaptability
fiveFull automation (autonomy)Unmanned operation without human controlSelf reroutingSelf-organization and autonomy
6Autonomous behavior in the group under any external conditions

On the basis of the examples given, one can already recognize a lot in common in classifications by levels and goals of autonomy. A generic classification that would successfully combine the various aspects together was developed and published by SafeTRANS as part of the technical planning for the implementation of highly automated systems. In it, the essential aspects of automation are divided into four classes:

  1. Functional automated systems can autonomously perform limited, well-defined tasks, such as automatic parking, landing, or fully automated production of one specific product. These systems cannot be trained during operation; cooperation with other systems is limited only by the exchange of contextual information, i.e. there is no collaboration.
  2. Mission-oriented systems can solve an unplanned sequence of familiar and achievable tasks, regardless of the situation. This may be, for example, autonomous driving a car on the highway or the operation of a single metro line. When implementing this, various optimization criteria may play a role, such as minimizing time or resource costs. Calculations for planning and optimization are performed by the system dynamically at the time of the task. These systems also can not be trained in the process; cooperation with other systems is limited to the exchange of information about the context of the task and about the system itself.
  3. Collaborative systems are such systems as robots, groups of vehicles or flying drones that move in a specific order or cooperate with each other to avoid collisions. In order to fulfill its mission, such systems are able to interact with other systems and people, as well as dynamically coordinate with each other their perceptions, interpretations, goals, plans and actions. These systems exchange relevant information with their partners, but, nevertheless, they are not trained.
  4. Self-sufficient (autopoietic) systems are systems that are able to independently develop their perceptions, interpretations, actions and abilities to work together, and also exchange these with other systems (including reproducing the behavior studied). These systems thus demonstrate human-like behavior and, to date, are still not implemented in practice. The ability to learn without control is the main characteristic of this class of systems.

An essential element of the presented classification is the ability to learn in autopoiesic systems. Today, cyber-physical systems cannot be endowed with such an ability, since there are no relevant regulatory requirements ensuring their reliable and safe operation, because it is impossible to prove the predictability and reliability of the system after its spontaneous change. Recent technological breakthroughs in the field of Deep Learning and high results in recognizing images and identifying patterns show, nevertheless, that developments in this direction and the possibilities of machine learning are developing at a very fast pace. However, there are still many obstacles along the way and more research is needed: neural networks can also develop themselves in an unintended direction or extract patterns from data that should not be recognized. Current research shows, for example, that the process of automatic learning also examines ethically undesirable historical data, such as sexual preferences or racist behavior. Thus, it is necessary to carry out appropriate control by ethical and legal norms. That is why the neural networks today can not be changed after the training phase. However, due to the large amount of input data and the complexity of neural networks, the problem of uncertainty in their behavior still remains. that the auto-learning process also examines ethically undesirable historical data, such as sexual preferences or racist behavior. Thus, it is necessary to carry out appropriate control by ethical and legal norms. That is why the neural networks today can not be changed after the training phase. However, due to the large amount of input data and the complexity of neural networks, the problem of uncertainty in their behavior still remains. that the auto-learning process also examines ethically undesirable historical data, such as sexual preferences or racist behavior. Thus, it is necessary to carry out appropriate control by ethical and legal norms. That is why the neural networks today can not be changed after the training phase. However, due to the large amount of input data and the complexity of neural networks, the problem of uncertainty in their behavior still remains.

In particular, in the Asian region, the role of autonomous systems is increasingly seen from the standpoint of their influence on humans. They should help people by simplifying their work. At the same time, a person is still a control element (human-in-the-loop), so in this case we can speak about cooperative intelligence. An example of this is the interaction of man and robot in the joint performance of a task. Such self-sufficient systems of the future transfer experience between machines and people and adapt their behavior. At the same time, there is an opportunity to address ethical issues. Behavior of machines in relation to a person is a field of computer ethics research. However, even simpler autonomous systems require an interface to a person. Relevant user interfaces are required to clearly inform

Autonomy requirements

This section presents the basic recommendations from SafeTRANS regarding safety, availability and the development of highly autonomous systems. The main difficulty for an autonomous system is the recognition of the surrounding reality.
The complexity of the processes of the world makes it impossible to carry out the multiple tests necessary to allow an autonomous system to operate. On this basis, it is recommended to implement an additional system of continuous monitoring of the system and train it on the basis of data obtained as a result of tests in real conditions. The diagram below shows such a meta-level learning process in which the system is tested in real conditions, and observational data after independent evaluation become the basis for the learning process.


SafeTRANS recommends the following activities for developing autonomous systems:

Action areaactivity
1. Models of the world1.1. Development of a common open industry standard for environmental models in various applications, in accordance with the stage of development and the level of complexity of the system. 1.2. Building a publicly available process and related infrastructure for building virtual systems testing. This requires: accredited institutions; public test environment; real-life testing specifications 1.3. Creation of arguments acceptable to permitting bodies and society, proving the safety of highly automated systems and based on the results of virtual verification and testing in real conditions.
2. Learning community2.1. Building an accessible process for learning based on real-world observations. For this you need: an independent accredited Trust Center; the obligation on the part of commercial organizations to voluntarily provide the authorized Trust Center with the necessary anonymous data; the transfer by the Trust Center of the analysis results back to the validation process.
3. Architecture3.1. Standardization of information exchange protocols between objects and situations in the industry to ensure their interaction with each other. 3.2. Standardized functional architecture for automated systems and their components, which confirms security and provides minimal functionality in simplified modes of use. 3.3. Uniform, consistent development process for highly automated systems, including the mandatory upgrade and upgrade capability. 3.4. An industry standard for certification / validation of the compatibility of system upgrades with an existing electrical architecture. 3.5. Clear standardized levels of degradation of systems with a guaranteed minimum of functionality
4. The guarantee of mutual compatibility of autonomous vehicles4.1. International classification of architectures of highly automated systems and their mutual compatibility. 4.2. The introduction of certificates of conformity for the architectures of the above classification, which are issued by an open public organization. 4.3. International, harmonized next-generation release process
5. Platform5.1. Providing a platform with basic services for autonomous vehicles at various stages of development. 5.2. The establishment of special standards for platforms that allow their independent certification. 5.3. Providing a software module that is able to show the current perceived state of the world at any given time, a picture of possible developments, as well as display, based on this, recommendations for action

Also popular now: