DAGR Enhances Goal-Conditioned Reinforcement Learning with State-Awareness

Traditional goal-conditioned reinforcement learning (RL) methods typically encode goals independently of the agent's current state, forcing the policy to infer the remaining actions. A new approach, DAGR, addresses this limitation by integrating state information directly into the goal representation. It uses a multi-scale gated cross-attention mechanism to dynamically adjust the goal embedding based on the current state. The DAGR framework builds upon existing late-fusion encoders, enhancing them with a state-conditioned perspective. While the method's name highlights "difference-aware" attention, experiments on OGBench navigation tasks showed improvements primarily due to a gated residual component, rather than the difference bias itself. For manipulation and puzzle tasks, DAGR performed comparably or slightly worse than baseline methods, indicating it's a specialized refinement rather than a universal enhancement.