Interview with Philipp Schwarz, TGW Software Services GmbH

Philipp Schwarz is a software developer at TGW.

TGW develops all essential components of its solutions itself: whether software, control, robotics or the mechatronic modules. We think beyond the limits of what is currently possible and break new ground. The vision of the autonomous fulfillment center inspires us. We shape the future with innovations and trend-setting technologies – and make them the high-performing standards in our industry. This claim distinguishes us and is constant motivation.

The TGW foundation company currently employs more than 3,700 people on three continents and generated sales of EUR 720 million in the 2018/19 financial year.

I work for TGW Software Services GmbH as a software developer. In addition tobeing employed by the an  OTH Regensburg, I am  employed as a studentical help.


  • Software development in the field of material flow
  • Work out and useof AI use cases to improve orderapproval inintralogistics plants.

OTH: Project for the  detection of poses by means of  CNNs / Research

TGW: Until an order leaves a  logisticcenter, various mechatronic components,also in connection with manual activities,, must be controlled  and efficiently coordinated. This  results in a large amount of data. This ranges  from sensor data from mechatronics to goods movements and order data. This large amount of data offers a high potential for the use of AI. One of the objectives that are pursued is, for example, Resresilience of logistics processes. Example topics  in which I am involved::

  • AI methods (e.g. genetic algorithms / reinforcement learning) to improve order processing and warehousing
  • Analysis of vehicle data of automatic shelving units

To my knowledge, there is no direct cooperation  between the TGW and the University of Regensburg, which goes beyond working students and thesis.

Economy:  AI is not the solution to all problems. Often simpler statistical models / classic algorithms are just as efficient,  but more stable and comprehensible. Students of AI-related programmes  (informatics, mathematics, media studies, …) should be able to give a sound reason for the reasons for which approach makes sense  or what advantages and disadvantages arise. This requires knowledge of problem modeling with and without AI methods. In the end, the added value for the company is crucial..

Social: AI techniques have and will continue to provide surprising applications (e.g. Deepfakes). In order to be able to assess the social implications of this “new”technology, a basic understanding is essential – regardless of the course of study.

TGW: As part of a university cooperation with JKU Linz, the use of reinforcement learning agents  for storage problems is being researched.

At the OTH I deal with the topic ofpose-estimation. Pose- I find  geometric deep learning very interesting, especially  generalizations  of CNNs on  Manigfaltikeiten.


  • Training IT specialists
  • Study of Mathematics (Master)
  • Software developer TGW / Research associate OTH Regensburg


AI knowledge:

  • Training:
    • Databases
    • Programming
    • Linux
  • Study:
    • Statistics
    • Schätztheorie, Statistical Learning (Tests, Regression, SVMs, Clustering, basic NNs, R)
    • optimization (e.g. Adam optimizer for NNs),
    • Robotics
  • Self-study / After (online courses, books,papers, conferences):):
    • CNN’s
    • Reinforcement Learning
    • Periphery  (Jupyter-Notebook, Tensorflow, Cuda, Docker  , …)
    • High-Dimensional Probability

Since “issues that have an AI-related relationship” is very broad, I cannot say yes or no to thisquestion. Due to my background in mathematics and computer science, I am able to understand most  concepts  quite quickly  and, if necessary, to familiarise myself more deeply with a topic. Nevertheless, you need a lot of sensitivity (data preparation, architecture optimization and domain knowledge) in order to be able to apply AI methods successfully. In short:: there is keinrecipe  to find out which approaches  / models  work. The can be frustrating,, but it also makes it interesting..

In addition, the productive use of AI requires much more than a good model. Powerful hardwareand softwareinfrastructurefor retrieving, storing and processing the data is just as  important. I.e. a wide range of specialists are needed for the successful implementation of AI projects,  who will hopefully be able to claim that they can find the topics in their arealeicht fallen“ easy. Generalists  like  in  scrumteams are ratherunwahprobably..

AI methods let the data speak for itself. Therefore,  facts given by this data are in the foreground. A good model will learn from these facts connections that might otherwise have been hidden..  Conversely,  by this data-centric view,,  incorrectly assumed correlations are revised. The focus is on what is actually and not on what couldbe. This is the added value of AI  (at least at the current stage). It is bought by  high calculation costs. Thee data-oriented view and  adaptive feedback via AI models I find very interesting. Intriguingly,    many algorithms  and approaches are inspired by processes outsidenature. In this context, I find Reinforcement Learning, the origin ofwhich can be traced back to  behavioural research,  particularly  interesting..

As mentioned above, the AI all-rounder does not exist. This should be reflected in the courses of study.

Mathematically oriented AI students are more likely to deal with topics such as optimization, modeling, statistics /  high–dimensional  probability.  The theoretical foundation is the focus.  Duringits computer science, AI students  also focused on hardware and software infrastructureuktur-topics. In between is  the data scientist. Different orientations must be possible. Also, the combination with other courses(e.g. Erstfach  Biology, …) interesting. Thus, one area couldprovide theanotforetics for the other.


AI / Statistics  lives  von the  application. Theory alone is not enough to get a sense of the different methods. Students need practical experience with models (NNs, RL, …) if necessary, have once mplemented and  testedrainiert and tested them yourself.  This requires appropriate hardware.  Practical projects with companies or other chairs are, in my opinion,  extremely conducive to applying theory.


The communicationwith experts from non-specialist fields is very important. For the successful application of AI methods, their knowledge is essential. The term AI often leads to falsepremise. Here, too, AI students should be able to communicate clearly what is feasible and what is not, or what youcan expect fromAI methods and  what  Aa u.a. wall they mean (see also 6.) .


The AI area has potential to bring application and research together in a cross-disciplinary  manner.

Furthermore, AI  kenntneaten studentscan gain importantadvantages. Nevertheless, the role of AI should not be overestimated.rden. Classical approaches and algorithms  must not be neglected..

Last edited on