Thinking Before Decision: Efficient Interactive Visual Navigation Based on Local Accessibility Prediction

Embodied AI has made prominent advances in interactive visual navigation tasks based on deep reinforcement learning. In the pursuit of higher success rates in navigation, previous work has typically focused on training embodied agents to push away interactable objects on the ground. However, such interactive visual navigation largely ignores the cost of interacting with the environment and interactions are sometimes counterproductive (e.g., push the obstacle but block the existing path). Considering these scenarios, we develop a efficient interactive visual navigation method. We propose Local Accessibility Prediction (LAP) Module to enable the agent to learn thinking about how the upcoming action will affect the environment and the navigation task before making a decision. Besides, we introduce the interaction penalty term to represent the cost of interacting with the environment. And different interaction penalties are imposed depending on the size of the obstacle pushed away. We introduce the average number of interactions as a new evaluation metric. Also, a two-stage training pipeline is employed to reach better learning performance. Our experiments in AI2-THOR environment show that our method outperforms the baseline in all evaluation metrics, achieving significant improvements in navigation performance.