This summer, a Palski intern team developed an environment for simulation of autonomous RPO’s using agents trained in a Reinforcement Learning framework. The environment allows for the simulation of the actions of multiple agents simultaneously, allowing new strategies for uncooperative RPO’s to arise organically. The current iteration of this project has been used to train several attacker and defender agents in a “capture-the-flag” scenario at GEO, allowing them to take actions to attempt to secure or disrupt flag capture. An associated visualizer tool was developed in order to facilitate analysis of simulated RPO’s.
To undertake this project, the intern team had to learn the theory behind Reinforcement Learning and autonomy as well as basic relative motion and RPO’s. Tools were developed using the Julia programming language, which also provided several open-source packages that were leveraged to support training for Reinforcement Learning. The team gained familiarity with a diverse set of concepts: everything from RPO’s to neural networks to development of frameworks for deep learning. They enjoyed their work, their time at the company, and the internship as a whole, and are excited to see what the future of this project holds.