Automation technology

Machines that perform an activity automatically have fascinated mankind since time immemorial. Even more so when their activities work intelligently - such as the so-called "Schachtürke," for example, which enthralled the Viennese court in the spring of 1770. At that time, the Austro-Hungarian court official and mechanic Wolfgang von Kempelen showed "Empress" Maria Theresa his latest work: a chess table at which sat a human-like, Turkish-robed figure that could move chess pieces with a mechanical arm - and also win against most opponents. For a long time, people puzzled over how the machine, which had clearly audible gears at work inside, managed to do this. The inventor kept silent all his life about the fact that it was simply a deception, a fraud: Hidden inside the apparatus was a human chess player who operated the mechanism. 
As technology advanced, however, it became more and more plausible that machines could actually achieve such capabilities. Thus, hardly anyone was surprised anymore when, in 1996, IBM's "Deep Blue" computer was able to beat a reigning world chess champion for the first time. In the intervening two centuries, humanity had witnessed the rise of technology and had become accustomed to the idea that machines could take over many a task previously reserved for humans, and could even do some things better. Calculating machines, for example, that tirelessly add up numbers without error. Or, later, industrial robots that assemble parts or process surfaces with a speed and/or precision unattainable by humans.

Whether such devices can be called "intelligent" is a matter of considerable debate. Today, at any rate, there are machines that can actually perform tasks autonomously and largely independently of humans. One example comes from an area that at first glance one would not consider all that innovative: In agriculture, autonomous working machines are already in use that - controlled by GPS and sensors - automatically pull their furrows in fields. Small, four-wheeled robots that work together as a fleet of autonomous special machines are also already ready for series production. These vehicles sow seeds, note where a crop is located, detect weeds with sensors (and analyze the data using AI methods) and kill unwanted plants with high-voltage pulses. Precision farming" is possible with such systems: Every field, every plant, receives exactly the treatment it needs - from irrigation and fertilization to pest control. Such farm robots still have their price, and many questions remain unanswered, for example about the use of the collected data. But they are predicted to have a bright future, not least because it is becoming increasingly difficult to find harvest workers and qualified personnel to work in the fields.

Another area of application for AI in material goods production is "predictive maintenance". Here, AI systems learn a connection between certain measurement data and the performance of machines. If there is a major deviation from the expected system behavior, this is automatically signaled by the system - then maintenance or repair is due. Real-time information about the condition of equipment allows potential production interruptions to be detected before they occur - and thus avoided. 
AI systems are also increasingly being used to ensure or increase the quality of industrial goods. Here, modern inspection systems play a major role, such as those being developed at the AIT Austrian Institute of Technology. In so-called "Inline Computational Imaging" (ICI), objects are moved past under high-speed cameras on a conveyor belt. Special algorithms have been developed for evaluating the camera data, which, among other things, create lightning-fast 3D reconstructions of the objects to be inspected, in which the smallest defects on the surface can be detected. Classic methods of image processing are used here, which are increasingly being supplemented with AI methods. This makes it possible, for example, to segment cracks in the test objects or to classify defects. In this way, the evaluation methods are successively refined and the quality of the inspection is continually increased.

A strategic research goal of the AIT is the development of autonomous working machines, such as excavators, cranes, forklifts, etc.. These are intended to support humans in their activities and take over heavy, dangerous or monotonous tasks. A test site was recently set up at the Seibersdorf site, for example, where an autonomous loading crane for tree trunks is being developed and tested. The task "Drive to the tree trunk, grab the tree trunk and bring it to the truck!" is a clearly defined and (with the appropriate equipment) easily solvable task for humans. For machines, however, this has hardly been possible until now. This is because there are many complex tasks and research questions behind the seemingly simple command. This includes, for example, the control of hydraulic components and the mechanical system, reliable task and motion planning including the localization of the machine's own position - even in changing environments - the correct grasping of objects, the robust perception of the environment, and object classification for the correct interpretation of the environment. AI systems help significantly with these complex tasks. 
The requirements for robots become even much greater when they leave "protected" and well-defined environments such as factory floors, fields or timber loading yards and have to find their way in the real world - with all kinds of disruptive and often unpredictable influencing factors that are impossible to take into account all in advance. Safety aspects are particularly problematic: Autonomous machines must never pose a danger to other objects and certainly not to humans. To ensure this, precise knowledge of their environment is necessary. In many applications, this point is the really "hard nut" to crack.

In many cases, it makes sense to create a "digital twin" of the environment. This can be done, for example, with the help of camera systems, radar sensors or laser measurements. Mathematical methods can be used to create a highly accurate 3D model of the environment, which is then segmented in a next step. Individual objects - such as a traffic sign or a pedestrian - are classified and assigned certain properties (e.g., a traffic sign has a fixed location, but pedestrians move around). Machine learning methods provide valuable services here, for example, they have learned to identify a traffic sign, link this information with others, and subsequently arrive at an understanding of the scene on the basis of which decisions can be made. 
In autonomous vehicles, this environment recognition is the basis for planning movements and controlling the car. According to Expert:inside, the correct interpretation of the environment model is the weakest link in autonomous driving. Incorrect classification of, for example, an oncoming vehicle or a crossing pedestrian can have disastrous consequences - as some famous accidents involving test cars from well-known companies have shown.

"The detection of entire situations is much more than the detection of a single sensory signal. In modern automation systems, different sensor modalities and imaging techniques are combined with data- and physics-based models. With the computing power available today, you can then use this information to optimize system behavior, save resources and make decisions about what to do next," explains Andreas Kugi, professor of complex dynamic systems at the Institute of Automation and Control Engineering (ACIN) at TU Wien and co-director of the AIT Center for Vision, Automation & Control. "AI plays a crucial role in this. The big challenge is how to interpret this data. To do this, you combine sensor fusion methods with machine learning and a priori knowledge in the form of physical models or semantic information."

A major problem with AI methods is that machine learning systems only ever deliver probability statements - and that is difficult to reconcile with safety-critical applications such as autonomous driving, where the highest possible level of safety must be demanded. Therefore, particularly high demands are placed on the testing and validation of such systems. Such methods are being developed, among others, in the "Dependable Systems Engineering" group of the AIT Center for Digital Safety and Security. In essence, the aim is to verify whether an AI is working correctly. To find out, an autonomous driving system is exposed to a critical situation with other cars or pedestrians in a simulation environment, for example, and can thus be tested under precisely defined conditions to determine whether the system in the vehicle reacts correctly. 
The more complex the systems to be tested become, the more the effort required for testing increases; the classic testing methods work less and less well in this context. Therefore, a second major research topic is the use of AI to improve verification and testing techniques. Important thrusts in further development are to make the procedures faster or to be able to better isolate and explain errors that occur. Optimizing such tests is also important in terms of balancing costs and benefits: In some cases, testing and verification account for 50 to 70 percent of the total development effort. 

Machines of the future are not intended to replace humans, but to support them. The goal here is to combine the respective strengths of humans and machines. Intelligent machines can relieve humans of strenuous, dangerous and monotonous work and make production processes more efficient, flexible, sustainable and resilient. Humans can concentrate on more complex, supervisory or creative activities and act in terms of holistic problem solving. The technical processes are thereby aligned with the needs and necessities of humans. 
The AIT Austrian Institute of Technology is hosting a discussion event with leading experts in this field on November 7, 2022 (2-4 p.m.) at the "Berlin Science Week" at the Einstein Center Digital Future (Wilhelmstraße 67): Andreas Kugi (TU Vienna, AIT), Lydia Kaiser (TU Berlin), Matthias Scheutz (Tufts University, AIT), Elisabeth André (University of Augsburg), Manfred Tscheligi (University of Salzburg, AIT), and Johannes Winter (L3S).
The panel discussion aims to provide concrete answers on how intelligent machines can help in addressing future challenges, such as resource conservation, climate protection, or labor scarcity. All interested parties are cordially invited to actively participate in this discussion.