Q-learning based multi-robot box-pushing with minimal switching of actions

Reinforcement learning has been commonly used in multi-robot decision making to cope with uncertainties in the environment. A shortcoming of this approach is the need for the robots to change their actions quite frequently, which is not feasible in a physical multi-robot system. This paper focuses on the development of a modified Q-learning algorithm with minimal switching of actions. By introducing the concept of reward threshold and changing the actions only when necessary, the new algorithm reduces the action switching probability effectively and improves the algorithm performance. A multi-robot box-pushing project is developed to validate the algorithm.