Robot learning can be improved by using a structured library of skills, modeling behaviors as graphs with bifurcating dynamics, and incorporating richer sensing. The presented work develops a model-based reinforcement learning approach using stochastic neural networks to learn forward models and stochastic Graph-DDP for planning. This achieves generalization of pouring skills over different materials through decomposition of dynamics into flow and amount, and demonstrates how tactile sensing can support manipulation. Overall, the use of structured representations, model-based learning, and multimodal sensing were shown to enhance robot skill acquisition and generalization.