CBAM-ST-GCN: An enhanced DRL-based end-to-end visual navigation framework for mobile robot

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Mobile robot navigation faces challenges in dynamic environments due to perception limits and unpredictable obstacles.
• Deep reinforcement learning (DRL) offers end-to-end solutions but struggles with high-dimensional inputs and non-stationarity, leading to unstable learning.
• Existing methods often require handcrafted rules, limiting adaptability.

Purpose of the Study:

• To propose an enhanced end-to-end visual navigation framework for mobile robots in dynamic environments.
• To improve the stability and convergence speed of deep reinforcement learning policies.
• To enhance collision avoidance capabilities in complex, obstacle-rich scenarios.

Main Methods:

• Introduced a Convolutional Block Attention Module (CBAM) to enhance visual perception by applying spatial and temporal attention.
• Designed a Spatio-Temporal Graph Convolutional Network (ST-GCN) to model the behavior of dynamic obstacles.
• Integrated a Velocity Obstacle (VO) method-based penalty into the reward function for improved collision avoidance.

Main Results:

• The proposed CBAM-ST-GCN framework demonstrated superior success rates in simulations compared to existing methods.
• Achieved significantly faster convergence speeds during policy learning.
• Real-world experiments confirmed the framework's effectiveness and adaptability in practical navigation tasks.

Conclusions:

• The CBAM-ST-GCN framework effectively addresses the challenges of visual navigation in dynamic environments.
• The integration of attention mechanisms and graph networks enhances perception and obstacle handling.
• The approach offers a robust and adaptable solution for real-world mobile robot applications.