Jan Stefan Zernickel is Head of Research and Development at InSystems Automation, Berlin. Currently he is involved in the ‘CrESt‘ compound project dealing with trends and questions of the future all around collaborating embedded systems.
AI supplies or promises systems that imitate resp. adapt human abilities, in the shape of learning systems. In view of such AI, where do you see the greatest potential for yourself resp. your company or research institute in the coming years? In short, what kind of application problems do you envisage to resolve in Berlin or plan to resolve with AI in the near future?
In connection with production and logistics, AI is above all very helpful in planning agents to solve transport problems. To this end we at InSystems are investigating in what way we can optimise the performance of logistical requirements by means of more intelligent systems on the one hand, and how to optimise these requirements that are adapted to a changing environment on the other hand.
We see the greatest potential in the fact that there is less effort required to plan a logistics system as this process can largely be automated. By using the customers’ data, our transport robots will be able to learn how to behave optimally by themselves without having a developer or manager to give explicit instructions. In doing this we hope to create more flexible fleets for different production surroundings on the one hand and to reduce the efforts in planning, implementing, and re-configuring logistics systems for autonomous vehicles on the other hand.
The idea is that our robots will not have to be adapted at all (or to a lesser extent) to every customer but turn into ‘Plug-and-Play’ systems.
Are we still talking about digitalisation or are you already going a step further towards an intelligent production? Can you give us examples from Berlin?
A largely digitalised operation for our vehicles is a prerequisite as the vehicle disposition connected to an ERP-system offers high self-sufficiency and in this way also provides the highest increase in performance and cost saving compared to the classic system. Also, many more advantages - such as being able to automatically trace production (with reference to process reliability, transparent material flow) without any interruptions - only become apparent in connection with a digitalised production.
We have a number of customers in Berlin who have already been using our transport robots over several years successfully. These are customers from the cosmetics-, printing-, or hygiene articles industry. Our logistics concepts fit very well into the processes of intelligent production and supply machines and workers, depending on the contract situation. We go beyond digitalisation and already take it for granted. We need production data to supply our robots with a context in which they can act.
The performance of AI-systems depends mainly on the amount and quality of data with which you are training these systems. Where do you see problems here and how do you solve them? How do you compare to other regions, and internationally, in respect of the availability of data?
Due to the nature of our products that are created individually for every customer, we have access to specific data sets that are made available to us by the relevant customer. We know what to expect of the environment and production time schedules. This is why we do not really have problems in obtaining data sets.
We do, however, see problems in the completeness of scenarios that reflect typical production data in increasingly more flexible factory surroundings. In order to be prepared for a deviation in a ‘standard’ production, we plan to reflect ‘corner cases‘ while developing our system strategies - meaning exceptional cases that may occur and which the system must be able to cope with. By doing this we are confident to intercept marginal conditions of situations that can occur compared to the planned standard operation. For example, peaks of production or deliveries that occur at an unusual time of day and which still have to be handled reliably by the fleet. We are furthermore thinking of training products on self-generated data sets. We are, however, not yet actively researching this area.
Time and again there are reports of AI-systems making incomprehensible decisions. It remains in the dark how learning is determined by neuronal nets. How do you face this problem?
Amongst others, we research within the CrESt (Collaborative Embedded Systems) project, in cooperation with BMBF and DLR, on how the system behaviour based on the design process can be developed predictably. So far, we do not yet use neuronal nets. In case of adaptive behaviour we do, however, use so-called constraints that should avoid mistakes and wrong optimisations. Constraints are programmed conditions that a system must stick to mandatorily.
However, the human being remains the superior, monitoring authority, who checks and evaluates the performance of the system and makes decisions on amendments.
Mr. Zernickel, thank you very much for the interview.