This paper contributes a novel framework to efficiently learn cost-to-go function representations for robotic tasks with latent modes. The proposed approach relies on the principle behind ensemble methods, where improved performance is obtained by aggregating a group of simple models, each of which can be efficiently learnedq. The maximum-entropy approximation is adopted as an effective initialization and the quality of this surrogate is guaranteed by a theoretical bound. Our approach also provides an alternative perspective to view the popular mixture of Gaussians under the framework of inverse optimal control. We further propose to enforce a dynamics on the model ensemble, using Kalman estimation to infer and modulate model modes. This allows robots to exploit the demonstration redundancy and to adapt to human interventions, especially in tasks where sensory observations are non-Markovian. The framework is demonstrated with a synthetic inverted pendulum example and online adaptation tasks, which include robotic handwriting and mail delivery.