Learning Individual Skills and Team Behaviors for Distributed Object Pushing

In this paper we discuss a learning approach to distributed object pushing. In the proposed approach, first the required individual skills for single-robot object pushing are learned using a fuzzy reinforcement learning method. Then, the robots learn how to coordinate their actions to push the object to the desired configuration cooperatively in a distributed manner. The proposed team-level learning benefits from the knowledge, which is in the form of a Q-table, that the agent has gained in its individual learning phase by a special design of reward signal and state-action representation. Each robot learns a threshold on its Q-value using a single state reinforcement learning method and pushes the object when the Q-value of its best action in the current state is above this threshold. The reward signal is designed based on the robots' Q-tables and no external critic is needed for learning cooperation. Simulation results show that the robots learn their individual skills and a cooperation protocol to push the object cooperatively