Simulation of multi-robot reinforcement learning for box-pushing problem

The box-pushing problem represents a challenging domain for the study of object manipulation in a multi-robot environment. Our box-pushing problem is based on the pusher-watcher approach, involving two pushers robots that learn the best strategy for cooperatively moving an oversized elongated box to a specified goal and one watcher robot acting as the environment. This paper presents a solution to the box-pushing problem based on reinforcement learning in a multi-agent system. Within the framework of the paper, a simulator has been developed to carry out the practical tests.