A fully distributed multi-robot navigation method without pre-allocating target positions

This study focuses on the multi-robot navigation problem with unpredictable state transition disturbance. The primary goal is to construct a fully distributed multi-robot navigation method without pre-allocating target positions. To this aim, a reinforcement learning based method is presented, in which a distribution of state transition module is proposed to guarantee adaptiveness when trained policies are applied in physical multi-robot systems. The method incorporates a centralized training but fully distributed execution framework. The former can eliminate non-stationarity of the environment, and the latter enables the robots to collaboratively handle partially observable scenarios. Mean while, the designed reward function can guide the robots to approach not pre-allocated target positions and the nearly optimal trajectories are achieved in continuous environment. After training, the robots make decisions independently, coordinate, and cooperate with each other to determine the next actions from their current positions before arriving in target positions without pre-allocation, in which the trajectories are nearly optimal with partial observation available for each robot. Simulations are performed with increasingly complex environments, such as the addition of static obstacles and randomly moving obstacles. The results show that the robots are able to achieve the primary goal with different state transition disturbance, which demonstrates the feasibility, effectiveness, and robustness. Furthermore, experiments are carried out using our multi-robot system corresponding to the simulation. The experimental results demonstrate the effectiveness and robustness of the proposed navigation method to handle a variety of typical robotic scenarios.