Language learning plays a crucial role in children's growth process. However, children's language learning is facing problems such as unbalanced distribution of educational resources and insufficient family support for language education. At the same time, the traditional teacher-centered teaching method lacks interactivity and fun, making it difficult to attract children's interest. For this reason, this study designs an AI Agent-based language learning robot for children. Throughout the study, we pay special attention to the second language (English) learning needs of non-native English-speaking children because English, as a globally accepted language, is of great significance for children's future development. This study relies on the natural language processing capability of the Doubao Big Model, integrates speech recognition and synthesis, multimodal sentiment analysis technology, and uses the Arduino hardware platform to realize embodied interaction. At the same time, we designed personalized teaching content that is suitable for different age groups and provides real-time learning progress tracking feedback. We invited 20 non-native English-speaking children in kindergartens to take the test, and through user interviews to understand their feelings and feedback on this new learning method. The results showed significant improvements in vocabulary growth, language expression, and self-confidence. The children felt that this design greatly increased their interest and engagement in learning English. In addition, the project will be a long-term experiment in the hope of providing more children with a quality learning experience.