Accurate prediction of foot pattern throughout various terrains is essential for achieving stable and adaptive control in powered lower-limb prostheses. In this study, a range of predictive models, including regression-based approaches, ensemble machine learning methods, and recurrent neural networks (RNNs), were systematically compared to determine their capability in generating oscillatory gait signals from tibial angular position while walking on varied terrain. The training and testing data were collected from ten healthy individuals while level-ground walking, ascending and descending stairs. Statistical models had poor accuracy and generalization, while ensemble methods (Gradient Boosting, Histogram-Based Gradient Boosting, and eXtreme Gradient Boosting) had moderate performance but remained sensitive to outliers and inter-subject variation. In contrast, recurrent architectures, i.e., long short-term memory networks (LSTM), had the highest predictive accuracy, with a mean correlation of 0.88, and a low root mean square error of 0.09 rad. Although the LSTM-based method provided slightly lower accuracy than some of the control methods inspired by the central pattern generators in the literature, it required only prosthesis-embedded sensors, avoiding the need for sensors on the intact or residual limb. Moreover, the compact network size and low computational load make it well-suited for embedded deployment, reducing cost, power consumption, and user burden. These findings place RNN-based approaches in line with ongoing research trends toward dynamic pattern generator–inspired controllers, offering robust, volitional-like control while maintaining practicality for real-world implementation.