Enhancing Text Game Agent Performance via Navigator

Behavioral cloning is a widely used approach for training text-based game agents. However, directly imitating low-level movement commands (e.g., move east) introduces ambiguous supervision, since the same movement command can lead to entirely different targets. This ambiguity lowers training efficiency and limits the agent's performance. To address this issue, we reformulate sequences of movement commands as high-level item navigation commands (e.g., navigate to red onion), which explicitly specify the intended goal. A parameter-free navigator module is then introduced to execute these commands by managing pathfinding and item tracking, allowing the agent to focus on higher-level decision-making. Experiments on the First Text World Problems (FTWP) dataset demonstrate that our method achieves a task completion rate of 96.8%, establishing a new state of the art and surpassing the previous best result of 91.2%.