The paper presents a curriculum-based approach for training RL agents on difficult planning task, such as Sokoban. The paper contributes a collection of approaches working together (from sub-task sampling, to MCTS as means to collect training experience) to solve complex tasks. The paper's contributions are novel and empirical evaluation shows strong results. The work is well positioned within the related works. Per the reviewers' comments and the authors' responses, in the camera ready version, the authors should: - Reframe the scope of the paper to around solving difficult Sokoban problems instead of general planning, - Discuss generalization to the other domains.