Multi-Object Navigation Using Potential Target Position Policy Function.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Embodied AI agents typically focus on single-object navigation.
• Real-world demands often involve sequential multi-object navigation, which current methods address inefficiently.
• Sequential execution of single-task methods leads to overlapping trajectories and reduced navigation efficiency.

Purpose of the Study:

• To propose an efficient reinforcement learning framework for multi-object navigation.
• To develop a hybrid policy that optimizes navigation by minimizing ineffective actions.
• To enable embodied AI agents to handle continuous, multiple user demands.

Main Methods:

• Utilizing a reinforcement learning framework with a hybrid policy.
• Embedding visual observations to detect and memorize semantic entities (objects).
• Projecting detected objects into semantic maps for long-term environmental memory.
• Implementing a hybrid policy with exploration and long-term planning strategies.
• Predicting potential target positions using semantic maps and prior knowledge of object relations.
• Planning paths to potential targets based on predicted positions.

Main Results:

• The proposed method demonstrates effectiveness in multi-object navigation tasks.
• Experimental results on Gibson and Matterport3D datasets show the method's generalization capabilities.
• The framework successfully reduces noneffective actions, improving navigation efficiency.

Conclusions:

• The developed hybrid policy framework enhances multi-object navigation efficiency in embodied AI.
• The integration of semantic maps and prior knowledge enables effective long-term planning and exploration.
• The method shows strong performance and generalization in large-scale 3D environments.