… ist der Titel des Roland-Mittermeir-Preises 2022 und wurde vom Förderverein Technische Fakultät mit EUR 1.500,00 ausgezeichnet. Der Autorin und Preisträgerin, Frau Dipl.-Ing.in Lorena Anna-Maria Gril, BSc wurde der Preis im Rahmen einer TEWI-Veranstaltung (gemeinsam mit der Vergabe der TEWI-Schüler*innenpreise und Best Performer Awards) am 22. September 2023 übergeben und die Arbeit wird hier kurz vorgestellt:
To ensure the safety of collaborative workspaces between humans and robots in industrial applications, the detection of dangerous situations that are not based on physical contact is essential. Therefore, the aim of the thesis, which was carried out in cooperation with Joanneum Research, is to develop a suitable prediction model that predicts potentially dangerous collisions between humans and robots. Since humans are the main source of uncertainty in such work environments, predictions about their future movements in the robotic system can be used to avoid physical contact. A common approach is to predict repetitive human movements using artificial neural networks, for example. In this work, a tensor-based approach is used to predict future movements based on the movement patterns of past seconds.
The data collection used the Optitrack system, which recorded x, y and z coordinates of 10 markers attached to the body during industrial assembly work. For example, the markers were attached to the hip, spine, shoulder, head, elbow, and hand. Two sets of data were recorded with different people. There are several challenges in modelling with the data obtained.
Firstly, the body has different characteristics that need to be taken into account. For example, the length between the hand and the elbow is fixed.
- Since people are not able to perform assembly operations in exactly the same way, all movement cycles are of different lengths.
- There is also a correlation between the individual joints. For example, to enable the movement of the hand, the elbow and shoulder must make certain movements. In addition, the data is highly dimensional.
- The biggest challenge is that the predictions have to be made in real time.
To consider the limitations of the human body (1), the data was transformed into joint angle space, where the lengths between the body parts are fixed when calculating back. To be able to perform modelling, a reference movement was created. The training data was stretched or compressed to the same length, then the median of the data at each time step was used to determine the reference cycle, thus also addressing the different lengths (2) of the movement recordings. For modelling purposes, four seconds were used along the reference movement to predict one second of movement. The properties of a tenor-based model allowed both the correlations between the joints and the high dimensionality (3) to be considered. Predictions could be made through a similarity measure that compares new movements with the reference movement. These predictions were calculated in real time (4), as the required computational operations were not complex.
The results depend on the parameters of the model, but a median overall error of prediction to the original data was found for all markers from 15-40 cm. In addition, the method was compared with the predictions of a neural network. On average, the presented method achieved a 50 cm lower total error.
In summary, the developed method addressed all challenges and achieved better results than the frequently used comparison method.