This paper addresses two key issues in deep reinforcement learning (DRL): lack of generalization to new target goals and data inefficiency. To tackle these challenges, the authors propose a target-driven visual navigation model and the AI2-THOR framework. The target-driven model learns a policy that embeds both the current state and the target goal, enabling it to generalize across targets and scenes without retraining. The AI2-THOR framework provides high-quality 3D scenes and a physics engine, allowing efficient collection of training samples. The method demonstrates faster convergence, better generalization, and real-world applicability with minimal fine-tuning. Experiments show that the model outperforms state-of-the-art DRL methods in terms of data efficiency and generalization, and can handle both discrete and continuous spaces. The paper also includes visualizations and robot experiments to validate the model's performance and adaptability.This paper addresses two key issues in deep reinforcement learning (DRL): lack of generalization to new target goals and data inefficiency. To tackle these challenges, the authors propose a target-driven visual navigation model and the AI2-THOR framework. The target-driven model learns a policy that embeds both the current state and the target goal, enabling it to generalize across targets and scenes without retraining. The AI2-THOR framework provides high-quality 3D scenes and a physics engine, allowing efficient collection of training samples. The method demonstrates faster convergence, better generalization, and real-world applicability with minimal fine-tuning. Experiments show that the model outperforms state-of-the-art DRL methods in terms of data efficiency and generalization, and can handle both discrete and continuous spaces. The paper also includes visualizations and robot experiments to validate the model's performance and adaptability.