The recent growth of robotic manipulation has resulted in the realization of increasingly complicated tasks, and various kinds of learning-based approaches for planning or control have been proposed. However, learning-based approaches which can be applied to multiple environments are still an active topic of research. In this study, we aim to realize tasks in a wide range of environments by extending conventional learning-based approaches with parameters which describe various dynamics explicitly and implicitly. We applied our proposed method to two state-of-the-art learning-based approaches: deep reinforcement learning and deep model predictive control, and realized two types of non-prehensile manipulation tasks: a cart pole and object pushing, the dynamics of which are difficult to model.