Dear RoboCup SPL Teams,

We (Team BadgerBots at UW Madison) are releasing our code that we used to compete in RoboCup 2023 standard platform league challenge shield. Our primary code and research contribution was the use of reinforcement learning (RL) to train our robot behavior for the attacker, defender, and goalie roles. Towards enabling other teams to use RL within their own behavior architectures, we are open sourcing both our robot code and the framework we used for training behaviors in simulation. The total code release aims to provide a basic working team using RL-derived behaviors and the code to train those behaviors.

We provide two repositories:

https://github.com/Badger-RL/AbstractSimRelease2023 AbstractSimRelease2023 is used for training neural network control policies for our main robot behaviors. The goalie policy’s training should work with the code provided. We are still applying finishing touches to make it easy to locally reproduce how we ran the attacker training process, and they will soon be pushed to this repository with instructions.

https://github.com/Badger-RL/BadgerRLSystemCodeRelease2023 BadgerRLSystemCodeRelease2023 is forked from BHuman’s 2021 code release and contains pre-trained neural network behaviors and provides a ready to use software suite for controlling a team of 5 Naov6 robots. It makes use of BHuman’s full software stack with the primary exception that main high-level behaviors are generated by neural network control policies.

We graciously thank BHuman for open-sourcing their codebase to advance the league.

Feel free to email b***s@wisc.edumailto:b***s@wisc.edu or create issues on either repository if you have questions or comments.

Sincerely,

BadgerBots