Self-Learning Agents Play Battlefield 1

In previous work we used reinforcement learning with concurrent actions combined with imitation learning to improve the training of self-learning agents in 3D environments.

To test the applicability of this methodology in a real-world setting (or at least in a real virtual setting) we initiated a collaboration with DICE with the goal to play the basics of Battlefield 1 using self-learning agents. The results of the collaboration are described in this video: