Less Repetition, Less Energy Cost: A Reinforcement Learning-Based Multiagent Energy-Saving Autonomous Exploration System.
TL;DR
This work proposes a distributed Multiagent Energy-saving Autonomous Exploration System (MEAES), and devise the consumption-exploration-balanced training framework (CEBF), which guides agents to transform from lazy exploration to energy-saving exploration strategies through dynamic reward shaping.
Abstract
Multiagent autonomous exploration in unknown environments is both meaningful and challenging. Due to the constraint of a partially observable environment, the collaboration among agents is often inadequate, leading to increased energy consumption. Worse still, a decrease in overall exploration performance may occur due to a single agent failure. To address these issues, we propose a distributed Multiagent Energy-saving Autonomous Exploration System (MEAES) based on reinforcement learning. To accurately evaluate the regional complexity of different branches and further enhance the long-term decision-making capabilities of agents, we introduce the dual-scale clustered observation (DSCO) module. The DSCO generates fine-grained representations based on graph modeling, enabling better characterization of both global and long-term exploration values. Furthermore, we propose an energy-saving action (EA) mechanism, which mitigates redundant exploration and reduces energy consumption by selective waiting actions and independent exploration strategies. Finally, we devise the consumption-exploration-balanced training framework (CEBF), which guides agents to transform from lazy exploration to energy-saving exploration strategies through dynamic reward shaping. Extensive experiments validate the effectiveness of MEAES, demonstrating effective zero-shot transfer performance across unseen environments.