A target-driven visual navigation method based on intrinsic motivation exploration and space topological cognition

Area of Science:

• Robotics
• Artificial Intelligence
• Cognitive Science

Background:

• Target-driven visual navigation is crucial for robotics.
• Existing methods often lack efficient exploration and environmental encoding.

Purpose of the Study:

• To propose a novel navigation architecture inspired by animal cognition.
• To enable simultaneous learning of exploration policy and environmental structure encoding.
• To improve robot navigation efficiency and adaptability.

Main Methods:

• Deep reinforcement learning framework with self-generated rewards (curiosity, count-based, temporal distance).
• Temporal distance for waypoint identification and episodic memory integration for environment structure.
• Integration of space topological cognition and a locomotion network for generalized navigation.

Main Results:

• The approach efficiently explores and encodes the environment in the DMlab 3D environment.
• Demonstrated improved capability in handling stochastic objects.
• Effective target reaching and detour behavior in navigation tasks, showing good environmental adaptability.

Conclusions:

• The proposed architecture offers a more generalized and adaptable approach to visual navigation.
• The integration of cognitive mechanisms enhances exploration and environmental understanding in robots.
• This work advances the field of target-driven visual navigation in robotics.