Emotion recognition is a research hotspot in the field of artificial intelligence. If the humanrobot interaction system can perceive human emotional behavior and express emotion, it will make the interaction between robot and human more natural. Humans acquire emotional information mainly through facial expression, semantic intonation and body language. Taking the NAO robot with high degree of freedom as an application platform, a humanrobot interaction system is designed for facial emotion recognition and body emotion expression. Firstly, the depthwise separable convolution algorithm is introduced to extract and classify features of facial expressions (e.g., angry, fear, sad, happy, surprise and neutral). Results showed that the prediction accuracy of FER2013 facial expression test set could reach 0711 by the trained network model. Secondly, the body movement of NAO robot are designed and classified according to six facial emotions. Finally, the realtime expression of the user′s emotional state by the robot is tested, and the feedback time is within 2 s. The statistical analysis of the prediction results of 10 consecutive frames is carried out.