This study employed a randomized between-subjects design to compare the effects of a job interview training simulation with the physical vs. virtual robot Furhat. The simulation was designed to assist students in preparing for their final exam and real-world job interviews. The system integrates Retrieval-Augmented Generation (RAG) and OpenAI's Large Language Model GPT4.o (LLM). 37 students participated in one of the two conditions: a physical or a virtual robot. In both conditions, enjoyment and attitudes toward robots significantly predicted students' willingness to reuse the training. Students, who had more positive attitudes towards robots, experienced greater advantages from the simulation. While male students had significantly more positive attitudes toward robots, female students experienced significantly higher levels of emotional tension during training. Despite these differences, both female and male students were equally willing to recommend and repeat the training. The results showed no significant differences between conditions in self-assessed learning outcomes, interaction quality, intrinsic motivation, or cognitive workload. These findings suggest that virtual robots may be a viable and scalable alternative to physical robots in educational settings. The study contributes to research on AI-powered physical and virtual pedagogical agents.