Artificial intelligence research

Festo brings together expertise in hardware and artificial intelligence to tackle unsolved problems

Festo is focusing on the smart production of the future. As a leader in technology and innovation for industrial automation, we aspire to establish artificial intelligence (AI) as a key technology and core competency and to make systematic use of this in providing our customers with automation technology solutions. We therefore carry out research into new opportunities and application areas.

Artificial intelligence involves a wide variety of methods and techniques, such as deep learning, reinforcement learning or bio-inspired AI. Many of these methods have been around for some time, but advances in processing power and modernised infrastructure mean that they can now be used widely in industrial production. AI enables us to tackle problems that were previously impossible to solve. For example, it is often not possible to work with model-based methods in the closed-loop control technology of complex systems with powerful fluid dynamics. The abstractions are too imprecise or too mathematically complex. Reinforcement learning will enable us to train such systems in future.

Reinforcement learning

Reinforcement learning enables machines to learn independently how to achieve a specified objective or solve a problem. The main benefit is that the computer finds a way itself, which may be quite different to the route that a human with learned experience would take. In many cases, this produces solutions that nobody had thought of before. The scope of application for this is wide, as reinforcement learning opens up vast potential, from control technology to robotics and supply chain planning.

Deep learning

Deep learning is particularly suited to individual skills that a robot needs to be very good at, such as gripping unfamiliar objects using the same gripper. Festo applies deep learning algorithms to robotics in the area of vision and also uses them to combine sensors for haptics, acoustics and infrared. This is necessary because robots have so far often operated based on cameras and stop working if the lighting fails, for example. Haptic, acoustic and infrared sensors make robots more robust and enable them to function in more difficult conditions, too.

Distributed inhomogeneous systems

In the field of distributed inhomogeneous systems, we are investigating whether it is possible for different systems to learn from each other: for example if a handling system can pass on its knowledge to a robot. In this case, passing on knowledge doesn't mean exchanging data, but rather that the systems communicate with each other and share their learned knowledge with each other. If this is possible, entire systems can optimise themselves and become better as more intelligent components are installed. For example, if a ball screw axis and an oversized cylinder are operating in succession, the cylinder notifies the ball screw axis that it needs to extend at a higher speed rather than at full pressure. This means that they save energy together and are more efficient.

Bio-inspired AI

Learning from nature is an important principle for Festo and one that extends beyond bionics. Nature can also serve as an example for algorithms. The structure of neural networks is partly modelled on the human brain, but the way that spiking neural networks function mimics the brain even more closely. Spiking neural networks can convey and process information independently of each other (and not just in layers like in neural networks). This means that they are more energy efficient and faster and could therefore be a highly promising approach for embedded systems with minimal processing power.

As part of the "Industry on Campus" collaboration with the University of Tübingen, we are investigating whether spiking neural networks could perform more complex or even creative tasks.

Generalisability and transferability

Alongside other joint research activities, the University of Tübingen and Festo are working on the generalisability and transferability of algorithms. To be able to transfer algorithms, it is important that we don’t need to train a separate model for every system and every application scenario.