HOZ++: Versatile Hierarchical Object-to-Zone Graph for Object Navigation.

Area of Science:

• Robotics
• Artificial Intelligence
• Cognitive Science

Background:

• Current deep learning agents struggle with object navigation in unseen environments.
• Human navigation relies on cognitive maps for recalling experiences and anticipating future events.
• Existing methods lack adaptability and robust decision-making in novel environments.

Purpose of the Study:

• To develop an agent navigation system inspired by human cognitive mapping.
• To improve object navigation performance in unseen environments.
• To create a versatile graph structure adaptable to various navigation methods and environments.

Main Methods:

• Proposed the Hierarchical Object-to-Zone (HOZ++) graph, modeling relationships between objects, zones, and scenes.
• Implemented a system where the HOZ++ graph guides agents by identifying zones and computing optimal paths.
• Enabled continuous updating of the HOZ++ graph with real-time observations for adaptability.
• Integrated the HOZ++ graph with existing reinforcement learning (RL) and modular navigation methods.

Main Results:

• The HOZ++ graph effectively guides agents to identify current and target zones, computing optimal paths.
• Continuous updates enhance the agent's adaptability to new, unseen environments.
• The proposed method demonstrated effectiveness and efficiency across four diverse simulators (AI2-THOR, RoboTHOR, Gibson, Matterport 3D) and a real-world environment.
• Successful integration with both end-to-end RL and modular approaches.

Conclusions:

• The HOZ++ graph significantly enhances object navigation capabilities by leveraging cognitive map principles.
• The method offers superior adaptability and decision-making compared to previous approaches.
• The HOZ++ graph presents a versatile and effective solution for complex visual navigation tasks.