Efficient Learning-Based Robotic Navigation Using Feature-Based RGB-D Pose Estimation and Topological Maps.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Robust indoor robot navigation often requires expensive sensors or large datasets.
• Existing methods face challenges with cost and data requirements.

Purpose of the Study:

• To propose a cost-effective RGB-D navigation pipeline.
• To enable reliable indoor robot navigation using feature-based pose estimation and a lightweight policy.

Main Methods:

• Coupling feature-based relative pose estimation with a lightweight multi-layer-perceptron (MLP) policy.
• Constructing a topological map using RGB-D keyframes and visual-geometric constraints.
• Utilizing LightGlue features and SVD for pose estimation and MLP for action prediction.
• Implementing real-time replanning using Dijkstra's algorithm upon detecting low visual similarity or obstacles.

Main Results:

• The system achieved reliable navigation across eight tasks in four environments, covering 190.44 m.
• The agent consistently stopped within 0.1 m of the goal.
• LightGlue demonstrated maximal information gain under lighting changes, validating its selection.
• The navigation pipeline operated effectively on commodity hardware.

Conclusions:

• The proposed modular design provides reliable navigation without metric SLAM or large-scale learning.
• The system is adaptable for future perception and policy enhancements.
• This approach offers a cost-effective solution for indoor robot navigation challenges.