["Developmental Science, Volume 29, Issue 4, July 2026. ", "\nABSTRACT\nIn this study, we applied machine learning tools to automatically track the positions of preschool children in a natural free play setting and derived spatial and temporal features from these data to identify social interactions between them. We observed a sample of 20 preschool children (10 female, 10 male; M ± SD = 3.95 ± 0.82 years) in groups of three children each. Friendship among children was assessed, and friend dyads were paired either with a mutual friend (n = 12 groups) or with a mutually disliked peer (n = 11 groups). We used a ceiling‐mounted camera to record 10‐min free play sessions of the 23 groups and used an automated keypoint tracking software to extract children's locations over time from the videos. From this data, we derived the following measures for each dyad within the group: distance, social orientation, and paired correlations of children's position and speed. Additionally, a human rater coded all occurrences of social interactions in the videos. Automated measures reliably predicted the occurrence of children's social interactions, validating our choice of spatial and temporal features. Friend dyads were closer, oriented more toward each other, and showed higher position and speed correlations than non‐friends. Social orientation and speed correlation varied over time, and speed correlation increased in mixed‐group contexts, especially among friends. These findings highlight the value of tracking‐based approaches for detecting both fine‐grained interactive behavior and affiliative ties, offering key insights into the spatial dynamics of young children's peer interactions.\n"]