Bio-inspired Decentralized Multi-robot Exploration in Unknown Environments

Multi-robot collaborative exploration is a fundamental precondition for a wide range of robot applications. However, in situations where physical communication networks cannot be deployed, existing methods encounter great challenge since collaboration mostly relies on the communication among robots. In this paper, a decentralized exploration approach for multi-robot systems is proposed to achieve efficient collaboration in communication-limited scenarios. Particularly, we present a novel multi-agent reinforcement learning approach called Multi-Agent Proximal Policy Optimization-Pheromone Interaction Mechanism (MAPPO-PIM) to realize self-adaptive and self-organized exploration. Inspired by nature swarms, we model the release and act process of pheromone and apply it into the motion of robots, which naturally forms an implicit interaction network based on the traces of pheromone. Moreover, we introduce the pheromone feedback into the reward shaping, which can be used to indirectly guide robots to explore different unknown areas without the requirements for explicit communication. We compare the performances of the proposed algorithm with several baseline exploration methods, and the experiment results indicate that our work obtains up to 45% collaboration improvement and 19% execution time reduction.