Less Repetition, Less Energy Cost: A Reinforcement Learning-Based Multiagent Energy-Saving Autonomous Exploration...

Area of Science:

• Robotics
• Artificial Intelligence
• Distributed Systems

Background:

• Multiagent exploration in unknown environments faces challenges like partial observability, inadequate collaboration, and increased energy consumption.
• Single agent failures can significantly degrade overall exploration performance.
• Existing systems often struggle with efficient long-term decision-making and energy management.

Purpose of the Study:

• To propose a novel distributed system, the Multiagent Energy-saving Autonomous Exploration System (MEAES), to address the limitations of current multiagent exploration methods.
• To enhance collaboration, reduce energy consumption, and improve robustness against agent failures.
• To enable efficient autonomous exploration in partially observable, unknown environments.

Main Methods:

• Developed a distributed Multiagent Energy-saving Autonomous Exploration System (MEAES) leveraging reinforcement learning.
• Introduced a dual-scale clustered observation (DSCO) module for accurate regional complexity evaluation and enhanced long-term decision-making using graph modeling.
• Implemented an energy-saving action (EA) mechanism with selective waiting and independent exploration strategies.
• Devised a consumption-exploration-balanced training framework (CEBF) with dynamic reward shaping to promote energy-efficient strategies.

Main Results:

• The proposed MEAES effectively manages energy consumption while maintaining high exploration performance.
• The DSCO module provides fine-grained representations, improving the characterization of global and long-term exploration values.
• The EA mechanism and CEBF framework successfully guide agents towards energy-saving exploration behaviors.
• Extensive experiments demonstrated the system's effectiveness and robust zero-shot transfer capabilities to unseen environments.

Conclusions:

• MEAES offers a significant advancement in multiagent autonomous exploration, particularly in energy-constrained and partially observable scenarios.
• The integration of DSCO, EA, and CEBF provides a comprehensive solution for efficient and robust exploration.
• The system exhibits strong adaptability and generalization, paving the way for more sophisticated autonomous exploration applications.