GNN enhanced reinforcement learning for robot navigation in complex topological networks.

Area of Science:

• Robotics and Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Intelligent robots face challenges in perceiving high-dimensional states and planning trajectories in complex environments.
• Existing methods struggle with the curse of dimensionality and static environmental representations.

Purpose of the Study:

• To develop an integrated Graph Neural Network-Reinforcement Learning (GNN-RL) framework for enhanced robot perception and adaptive trajectory planning.
• To address limitations in high-dimensional environmental state perception and decision-making for robots in complex topological settings.

Main Methods:

• Utilized Graph Neural Networks (GNNs) to model environmental topology, abstracting entities into nodes and encoding spatial/semantic relationships as edge features.
• Implemented a Soft Actor-Critic (SAC) algorithm for continuous control, integrating GNN-extracted topological features into a reinforcement learning agent.
• Employed a multi-objective reward function (safety, progress, smoothness) and end-to-end joint optimization of GNN and RL parameters.

Main Results:

• The GNN-RL framework effectively compresses high-dimensional environmental information into low-dimensional embeddings, enabling efficient action selection.
• Demonstrated a favorable balance between perception accuracy and decision-making efficiency compared to DQN, PPO, and A* algorithms in simulations.
• Achieved reliable and adaptive robot navigation and trajectory planning in structured, dynamic environments.

Conclusions:

• The proposed GNN-RL approach offers a robust solution for intelligent robots navigating complex environments.
• Integrating GNNs with RL overcomes limitations of static graph learning and enhances adaptive decision-making capabilities.
• The framework provides a promising direction for improving robot autonomy and performance in real-world applications.