Sports action recognition (SAR) has been widely applied in athletic training and sports injury rehabilitation. Compared with generic actions, sports actions are typically faster, more complex, and exhibit smaller inter-class variations. Although deep neural network–based methods have achieved strong performance on generic action recognition, their limited exploitation of fine-grained cues and insufficient interpretability hinder their applicability to SAR. Genetic programming (GP), with its potential interpretability and strong learning capability, offers a promising solution for SAR. However, GP has been rarely explored in SAR, and its potential remains largely unexplored. This work first designs a GP-based multi-layer feature extraction framework to investigate the feasibility and potential of GP for 2D skeleton-based SAR. It then proposes a multi-tree GP method for 3D skeleton-based action recognition to further improve GP performance in SAR. In addition, a multi-tree GP approach is developed for multimodal SAR. Finally, for group SAR, we introduce a new multi-person–object skeleton graph representation and a hybrid method that integrates GP with graph convolutional networks.