Solving the Puppet Cube optimally presents a significant challenge for classical algorithms due to its complex branching factor. Inspired by the DeepCubeA methodology, this project implements a hybrid solver that leverages Reinforcement Learning to overcome the limitations of manual heuristics. We document the transition from a pure A* baseline to a learned heuristic model, demonstrating how data-driven search strategies can efficiently solve shape-shifting puzzles that were previously considered computationally expensive.