Human-technology collaboration relies on verbal and non-verbal communication. Machines must be able to detect and understand the movements of humans to facilitate non-verbal communication. In this article, we introduce ongoing research on human activity recognition in intralogistics, and show how it can be applied in industrial settings. We show how semantic attributes can be used to describe human activities flexibly and how context informantion increases the performance of classifiers to recognise them automatically. Beyond that, we present a concept based on a cyber-physical twin that can reduce the effort and time necessary to create a training dataset for human activity recognition. In the future, it will be possible to train a classifier solely with realistic simulation data, while maintaining or even increasing the classification performance.