This thesis explores the development of intelligent agent policies using reinforcement learning, for playing table football against human opponents on a real automated table football table. First, a simulator was developed, whose purpose was both training and testing of policies. For the purposes of simulation-to-real transfer, identification of player rods was performed, where measurements of real rod responses were used to fit simulation parameters. Other discrepancies were compensated during training of individual policy. For better generalization against human opponents, an artificial opponent was developed, which was active during the training process of several policies. The artificial opponent is based on state-machine logic. Three policies were developed using reinforcement learning: a goalkeeper policy, a defense policy and an attacker policy. The midfield rod was made to use the same state-machine logic as the opponent. The developed policies were trained only in simulation, without additional training in the real environment. First, the policies were individually tested in simulation, where we evaluated their ability to defend their own goal (goalkeeper and defense policies) and to score into the opponent's goal (attacker policy). In the end, all policies, with the addition of the midfield state-machine agent, were evaluated against several human opponent pairs in the real environment.
|
