Multi-Agent Reinforcement Learning

Host: Automatic Control Laboratory (IfA)

Semester Project / Master's Thesis

Robot soccer is a team sport: every player has to decide where to move, when to go for the ball and when to support or defend, based on its own noisy view of the game and limited communication with its teammates. Hand-crafted behaviors struggle to capture this kind of coordination.

In this project, you will use multi-agent reinforcement learning to train cooperative team strategies in simulation and transfer them to our humanoid robots. The focus is on learning decentralized policies that each robot can run on its own, while still producing coordinated play such as positioning, passing and defending as a team.