Visual Target-Driven Robot Crowd Navigation with Limited FOV Using Self-Attention Enhanced Deep Reinforcement...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Simultaneous Localization and Mapping (SLAM) methods face challenges in dynamic, unpredictable environments.
• Mobile robot navigation in crowded spaces requires robust obstacle avoidance and target pursuit capabilities.

Purpose of the Study:

• To develop a visual target-driven navigation method for mobile robots using self-attention enhanced deep reinforcement learning (DRL).
• To overcome limitations of traditional SLAM in dynamic environments and improve robot navigation efficiency.

Main Methods:

• Utilized Twin-Delayed Deep Deterministic Policy Gradient (TD3) algorithm for navigation policy development.
• Employed a single RGB-D camera and convolutional neural network (CNN) for environmental feature extraction.
• Integrated a self-attention network (SAN) to compensate for limited field of view (FOV) and aid target re-acquisition.

Main Results:

• Achieved a higher success rate in dynamic environments compared to traditional methods.
• Demonstrated a shorter average target-reaching time.
• Validated the method's effectiveness in enhancing target search with limited FOV.

Conclusions:

• The proposed self-attention enhanced DRL method offers a robust solution for mobile robot navigation in challenging, dynamic environments.
• The system provides hardware simplicity, cost-effectiveness, and ease of real-world deployment.
• This approach significantly improves robot's ability to navigate and reach targets efficiently, even when temporarily lost.