This paper discusses research efforts on social robots that recognize interpersonal relationships. These investigations are carried out by observing group behaviors while the robot interacts with people. Our humanoid robot interacts with children by speaking and making various gestures. It identifies individual children by using a wireless tag system, which helps to promote interaction such as the robot calling a child by name. Accordingly, the robot is capable of interacting with many children, causing spontaneous group behavior from the children around it. Here, group behavior is associated with social relationships among the children themselves. For example, a child may be accompanied by his or her friends and then play together with them. The hypothesis is proposed that the interactive robot prompts a child’s friends to accompany him or her; thus, it is possible to estimate their friendship by simply observing their accompanying behaviors. To verify this hypothesis, a field experiment for two weeks in a Japanese elementary school was conducted. In the experiment, two “Robovie” robots were placed where children could freely interact with them during recesses. As a result, it was found that they mostly prompted friend-accompanying behavior. Moreover, it was possible to estimate some of their friendly relationships, in particular among the children who often appeared around the robot. For example, 5% of all friendships could be estimated with 80% accuracy, and 15% of them with nearly 50% accuracy. Thus, this result basically supports the hypothesis on friendship estimation from an interactive humanoid robot. It is believed that this ability to estimate human relationships is essential for robots to behave socially.