Object detection method for improving the generalisation of deep reinforcement learning

One key reason for the gap between deep reinforcement learning goals and applications is that trained intelligences overfit local training data features, and models trained with perception-decision in a single network are sensitive to small environmental changes. This restricts the agents from learning real rules from experience. In this paper, we use a target detection model to recognise objects in a visual scene, learn perceptual level knowledge and use the generalisation ability of the target detection model to reduce the observation overfitting of the agent and improve the robustness of the agent in vision. After the object instances are extracted by the target detection model, we use a model structure equipped with a relational inference mechanism to model the relationships between the objects, which further improves the model's ability to generalise to unseen scenes by introducing a relational inductive bias.