ST-GICM: A Spatiotemporal Graph Learning Framework with Intrinsic Curiosity for Robust Autonomous Exploration.

Area of Science:

• Artificial Intelligence
• Robotics
• Machine Learning

Background:

• Deep reinforcement learning (DRL) and graph neural networks (GNNs) have advanced autonomous exploration.
• Existing methods face challenges in long-horizon decision-making and sustained exploration under partial observability and sparse rewards.

Purpose of the Study:

• To propose a spatiotemporal graph learning framework (ST-GICM) to enhance robustness and efficiency in autonomous exploration.
• To address limitations in current graph-based exploration methods for challenging environments.

Main Methods:

• Developed ST-GICM, integrating graph-structured encoding, temporal memory, and intrinsic curiosity.
• Employed Graph Attention Network (GAT) and Spatiotemporal Reasoning Core (STRC) for dynamic graph encoding and temporal fusion.
• Designed an Intrinsic Prediction Module (IPM) to generate intrinsic rewards based on prediction error for sustained exploration.

Main Results:

• ST-GICM demonstrated superior performance in coverage rate, success rate, and reduced oscillation count compared to baselines.
• The method maintained comparable trajectory costs.
• Achieved significant improvements in complex, procedurally generated topological environments.

Conclusions:

• ST-GICM effectively improves autonomous exploration in partially observable, sparse-reward environments.
• The framework's integration of graph learning, temporal memory, and intrinsic curiosity enhances robustness and efficiency.
• Outperforms existing methods, showcasing its superiority for challenging exploration tasks.