Field of Research: Artificial Intelligence

Festo Combines Hardware and AI Expertise to Tackle Previously Unresolved Issues

Festo is positioning itself for the intelligent production of the future. In our role as a technology and innovation leader in industrial automation, we aim to establish artificial intelligence (AI) as a key technology and core competence and to systematically use it for our customers’ automation solutions. As a result, we are currently exploring new possibilities and areas of application.

Artificial intelligence encompasses a wide range of methods and techniques. These include methods such as deep learning, reinforcement learning, or bio-inspired AI. Many of these methods are not new, but increased computing power and a modernized infrastructure now make their widespread use in industrial production possible. AI allows us to tackle problems that were previously impossible to solve – in the control engineering of complex systems with strong fluid dynamics, for example, working with model-based methods is often not possible. The abstractions are too imprecise or mathematically complex. Reinforcement learning will allow us to train such systems in the future.

Reinforcement Learning

Reinforcement learning allows machines to learn how to achieve a given goal or solve a problem on their own. The major advantage is that the computer itself finds a path that may be quite different from the one a human with learned experience would take. In many cases, this generates solutions that no one thought of before. Reinforcement learning holds tremendous potential for applications ranging from control engineering and robotics to supply chain planning.

Deep Learning

Deep learning is particularly suitable for individual skills that a robot must be able to do extremely well, for example grasping unknown objects, but always using the same gripper. Festo transfers deep learning algorithms in the field of vision, but also in order to integrate haptic, acoustic, and infrared sensors, to the robot technology. This is because up to now, robots have often been camera-based and cannot continue working if the lights go out, for example. Haptic, acoustic, and infrared sensors make robots more robust and function even under more difficult conditions.

Distributed Inhomogeneous Systems

In the field of distributed inhomogeneous systems, we are researching whether it is possible for different systems to learn from each other – whether, for example, a handling system can pass on its knowledge to a robot. Knowledge in this case does not mean that data is exchanged, but that systems communicate with each other and share their learned knowledge. If this is possible, entire systems can optimize themselves and become better the larger the number of intelligent components installed. If, for example, a ball screw axis and an oversized cylinder are working in tandem, the cylinder will tell the ball screw axis to extend at a higher speed instead of at full pressure. In this way, they save energy and work together more efficiently.

Bio-Inspired AI

Learning from nature is an important principle at Festo, and not only in the field of bionics. Nature can also be a model for algorithms. Neural networks are partially modeled on the human brain in terms of their structure, but the way spiking neural networks work is even closer to the human brain. They can transmit and process information independently (and not only layer by layer, as is the case with neural networks). As a result, they use energy more efficiently and operate faster, meaning they could serve as a promising approach for embedded systems with little computing power.

Within the framework of the “Industry on Campus” project with the University of Tübingen, we are jointly researching whether spiking neural networks can solve more complex or even creative tasks.

Generalizability and Transferability

Among other joint research activities, the University of Tübingen and Festo are working on the generalizability and transferability of algorithms. It is important that a separate model does not have to be trained for each system and use case so that we can transfer algorithms.