We propose a novel planning technique for satisfying tasks specified in\ntemporal logic in partially revealed environments. We define high-level actions\nderived from the environment and the given task itself, and estimate how each\naction contributes to progress towards completing the task. As the map is\nrevealed, we estimate the cost and probability of success of each action from\nimages and an encoding of that action using a trained neural network. These\nestimates guide search for the minimum-expected-cost plan within our model. Our\nlearned model is structured to generalize across environments and task\nspecifications without requiring retraining. We demonstrate an improvement in\ntotal cost in both simulated and real-world experiments compared to a\nheuristic-driven baseline.\n