StructRL: Model-Free Reinforcement Learning with Replay Graphs

StructRL turns reinforcement learning experience into an explicit replay graph. For DQN on three sparse-reward MiniGrid tasks, final success rose from 0.10 to 0.90 for DoorKey, 0.10 to 0.80 for KeyCorridor, and 0 to 0.50 for ObstructedMaze. For SAC on continuous PointMaze, graph-derived direction labels raised final success from 0.17 to 1.0 on seven matched runs.