Mutli-UAV target-finding in simulated indoor environments using deep reinforcement learning

Searching indoor environments in the presence of unknown obstacles with multiple UAV agents remains a challenge. This paper presents a framework for target-finding using a combination of traditional POMDP based planning and Deep Reinforcement Learning. The implementation breaks the problem into two separate stages of planning and control, with both stages modelled as a Partially Observable Markov Decision Process (POMDP). Global decentralised planning is provided using a modern online POMDP solver, while a modern Deep Reinforcement Learning algorithm is used to provide a policy for local control. Our results indicate that such a framework is capable of target-finding within a simulated indoor test environment in the presence of unknown obstacles, and once extended to real-world operation could enable UAVs to be applied in an increasing number of applications.