Synthetic environments with verifiable rewards enable RL training for game-playing agents far more effectively than demonstrations alone—Player-1-9B reaches 38% success on MineDojo compared to 13% for baselines, showing that programmatic environments can scale agent learning beyond human-collec...
Player-1 is a framework that teaches AI agents to play games by combining supervised learning on gameplay videos with reinforcement learning in synthetic game-like environments. The approach extends typical computer-control models with new action types and trains on 16 custom environments, achieving significant improvements on game benchmarks like MineDojo and ViZDoom.