Reinforcement learning agents in complex tasks often require extensive exploration of large state spaces before useful structure emerges. Humans, in contrast, leverage high-level semantic knowledge such as recipes, instructions, or labels, and adapt to new tasks by grounding that prior knowledge in the environment. This thesis develops Knowledge Graph-Augmented Reinforcement Learning (KG-RL), a method built to augment RL policies with structured graph information. This work demonstrates the ability of our method to inject this structured prior into a variety of policy networks, and to act across a range of environments. The adapter consumes a task knowledge graph (KG) merged each step with a scene graph (SG), runs a Graph Convolutional Recurrent Network over the merged graph, and pools the result through a small recommender into a fixed-width feature vector concatenated with the backbone's observation features before the heads. The backbone itself is untouched, so the adapter slots into learned policies such as CNN--MLP, SoftMoE--LSTM, GTrXL, and PoliFormer without modification; bringing it to a new environment requires only enumerating a handful of entities and relation templates before training.

The KGs are built from a small set of named entities (objects, tools, room types, achievements) plus three to six relation templates: instantiated against simulator-exposed tables in symbolic environments, or mined by a one-time pre-pass of the same perception pipeline that builds per-step SGs in open-world AI-Habitat. The SG is constructed per step either algorithmically (geometric probes over the simulator state) or perceptually (open-vocabulary detection and promptable segmentation over the agent's RGB frame). Across four environments---Overcooked-AI, MiniGrid, Craftax, and AI-Habitat ObjectNav---and the four backbones, the adapter delivers two improvements independent of backbone choice: \emph{faster training to the same final policy} (single-agent Overcooked, MiniGrid), reaching the same asymptotic reward in up to $\sim$$3\times$ fewer environment steps in certain environments, and \emph{higher final policy under a fixed budget} (two-agent Overcooked, Craftax, AI-Habitat), where on Craftax and AI-Habitat our method overtakes the available published baselines. Both gains scale with task complexity, and the adapter is robust to substantial KG corruption, retaining its advantage with half of the graph's nodes removed.