A Framework for Studying Reinforcement Learning and Sim-to-Real in Robot Soccer

In Very Compact Dimension Soccer two groups of three robots contend to score aims in opposition to every other. The behaviour of robots is usually programmed for every scenario. Reinforcement finding out could be employed to strengthen the abilities of robots nevertheless, serious-planet teaching is impractical since of the degradation of components and the usage of energy and time.

Very Compact Dimension robotic soccer level of competition. Image credit rating: Hansenclever F. Bassani, Renie A. Delgado, José Nilton de O. Lima Junior, Heitor R. Medeiros, Pedro H. M. Braga, Mateus G. Machado, Lucas H. C. Santos, Alain Tapp, arXiv:2008.12624

A the latest review proposes a framework for sim-to-serious teaching. In this case, robots are experienced in simulations and the acquired plan is transferred to the serious planet. It is shown that this tactic qualified prospects to a broader repertoire of behaviours than human-intended plan, but strikes are slower and much less precise. The success of reinforcement finding out was evaluated in the 2019 Latin American Robotics Levels of competition. Here, it was a initially time a team of robots experienced by the reinforcement finding out has received in opposition to groups which operated by human-intended guidelines.

This short article introduces an open up framework, called VSSS-RL, for finding out Reinforcement Finding out (RL) and sim-to-serious in robot soccer, focusing on the IEEE Very Compact Dimension Soccer (VSSS) league. We suggest a simulated ecosystem in which ongoing or discrete handle guidelines can be experienced to handle the total habits of soccer brokers and a sim-to-serious technique centered on area adaptation to adapt the attained guidelines to serious robots. Our benefits show that the experienced guidelines acquired a wide repertoire of behaviors that are complicated to put into practice with handcrafted handle guidelines. With VSSS-RL, we were being in a position to conquer human-intended guidelines in the 2019 Latin American Robotics Levels of competition (LARC), achieving 4th spot out of 21 groups, remaining the initially to apply Reinforcement Finding out (RL) efficiently in this level of competition. The two ecosystem and components requirements are out there open up-resource to enable reproducibility of our benefits and additional scientific studies.

Hyperlink: https://arxiv.org/abs/2008.12624