Deep Reinforcement Learning-Based Automatic Exploration for Navigation in Unknown Environment.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Automatic exploration is crucial for applying robotic systems to social tasks in unknown environments.
• Traditional rule-based methods struggle with environmental variability and sensor differences.
• Learning-based control methods offer adaptability but suffer from low efficiency and poor simulation-to-reality transfer.

Purpose of the Study:

• To develop a general and modular framework for automatic robotic exploration.
• To propose a deep reinforcement learning (DRL) based decision algorithm for enhanced exploration strategies.
• To improve learning efficiency and adaptability in unknown environments.

Main Methods:

• Decomposition of the exploration process into decision, planning, and mapping modules.
• Development of a DRL algorithm utilizing a deep neural network to learn exploration strategies from partial maps.
• Experimental validation on physical robots to assess simulation-to-reality transferability.

Main Results:

• The proposed DRL algorithm demonstrates superior learning efficiency and adaptability in unknown environments.
• Experimental results confirm the algorithm's effectiveness in diverse and unpredictable settings.
• The learned exploration policy successfully transfers from simulation to a physical robot.

Conclusions:

• The modular exploration framework enhances robotic system adaptability.
• The DRL-based decision algorithm offers an efficient and transferable solution for autonomous robotic exploration.
• This approach advances the practical application of robots in complex, real-world scenarios.