中国科学院自动化研究所机构知识库

Multi-target coverage with connectivity maintenance at scale
remains challenging in current research. A novel double-observation pol-
icy learning framework (DOPLF) aiming at multi-agent system deploy-
ment in large scale coverage problems is proposed in this work. DOPLF
introduces observations from both global and local perspectives to en-
courage more ecient exploration in large scale coverage scenarios with
massive state spaces. Specifically, the local-level observation is derived
from target partition to provide regional target density for each agent,
and global-level observation provides an overall information of the en-
vironment. Both observations are then fed into the subsequent learn-
ing modules that primarily adopt graph attention network and proximal
policy optimization based reinforcement learning algorithm to generate
a distributed policy. Further, curriculum learning is applied to enhance
the model adaptability in larger team sizes. Finally, the proposed method
outperforms the baseline method in coverage rate and training eciency
in simulations with the number of targets ranging from 50 to 500.