This paper proposes a novel approach to natural language understanding (NLU) in service robots called RoboNLU, which leverages pre-trained language models along with specialized classifiers to extract meaning from user commands. Specifically, the proposed system utilizes the Bidirectional Encoder Representations from Transformers (BERT) model in conjunction with slot, intent, and pronoun resolution classifiers. The model was trained on a newly created, large-scale, and high-quality GPSR (General Purpose Service Robot) command dataset, yielding impressive results in intent classification, slot filling, and pronoun resolution tasks while also demonstrating robustness in out-of-vocabulary scenarios. Furthermore, the system was optimized for real-time processing on a service robot by leveraging smaller, quantized versions of the BERT-base model and deploying the system using the ONNXruntime framework (Code and data available at https://github.com/RoboBreizh-RoboCup-Home/RoboNLU).