A Python and TensorFlow-based machine learning project attempting to build a model to predict the solvability of a Sokoban level. The motivation was to test the capabilities of CNN-based models to solve a longer horizon planning task, and also to simply develop an end-to-end machine learning application. The dataset was compiled from various collections of Sokoban levels online, with their pre-processing and cleaning done by me. I used a fairly standard CNN-based model that’s also been used for image recognition tasks. The model is able to achieve an accuracy of $74.69%$, which is marginally better than the majority label rate of $72.5%$ in the dataset.