Natural language encodes rich sequential and contextual information. A task plan for robots can be extracted from natural language instruction through semantic understanding. This information includes sequential actions, target objects and descriptions of working environment. Current systems focus on single-domain understanding such as household or industrial assembly settings, and many rule-based approach have been developed in this context. Thanks to the development of deep learning, data-driven contextual language understanding shows promising results. In this work, an information extraction system is proposed for domain-independent understanding of robotic task plans. The developed approach is based on a pre-trained BERT-model and a syntactic dependency parser. To evaluate the performance, experiments are conducted on three different datasets.