project

Today’s production lines are heavily automated and run with high throughputs. The later in
the processing step the better these lines can be optimized to specific properties of the
material or product to be further processed. Product handling/processing is usually set up as
a sequence of individual production line steps with each step being highly optimized for its
specific processing task. This is only possible through precise definition of the input and the
output of each step in the entire production line. As the precision and the performance of the
machines increase, the output (i.e. input of the next step) can be defined within more narrow
bands which leads – assuming no (undetected) errors – to further optimization potential.
For non-deterministic inputs, today’s approaches still struggle in finding sufficiently versatile
processing approaches to guarantee a certain output quality for the subsequent step. Today,
such steps are usually performed by specifically trained human personnel. Despite the
necessary (safety) training, these steps have highly repetitive patterns (are dull). They
usually occur at the very end of the production line where detailed/customized finishings
have to be performed, or, more often, at the beginning where raw materials have to be
shaped into an initial condition. Steps in the beginning often are performed outdoors in harsh
environmental conditions (are dirty), and comprise of significant potential of accidents and
dangerous situations (are dangerous).
In this project, we specifically tackle the DDD (Dull, Dirty, Dangerous) tasks for humans by
automating the challenging (first) steps where (biological) raw material is handed over to the
production line. We split the task into three main investigation areas: i) understanding of the
material geometry in 3D and subsequent handling strategies, ii) automated navigation and
control of the existing machinery (previously controlled by hand) through visual-inertial based
closed loop control and through iii) retrofitting the existing machinery with appropriate novel
autarkic sensors capable of providing feedback to both the control and machine learning
algorithms.
The retrofitting aspect of the package (AI based 3D understanding, control and navigation,
autarkic sensors) will allow usage of the approach even for small companies/lots in variety of
different (existing!) production lines and industry areas. In addition, through the versatility of
the approach (e.g. automated retraining of the AI network through the sensor’s feedback) it
will allow unprecedented flexibility, performance, and dynamic adaptation capability of a
production line to increase the number per product, but also amount of different products.
The proposed approach will yield a deterministic output already after the first processing step
eliminating human based variances which will allow further optimization for (all) following
steps leading to increased end-product quality.