Player-1: Learning to Play from Synthetic Experience
Player-1-9B reached 38.33% success on MineDojo-300, up 27.8 percentage points from its supervised starting point.
The researchers first fine-tuned computer-use models on 6,044 curated gameplay trajectories, then trained them with reinforcement learning in sixteen synthetic 2D and 3D environments. The framework adds relative pointer motion, sustained inputs and combined keyboard-and-mouse actions. The larger Player-1-35B-A3B gained 17.7 percentage points on MineDojo-300, and both models improved over supervised training on ViZDoom-10 and GameWorld. Broader coverage of synthetic environments improved transfer beyond the reinforcement-learning training suite.
HF Daily Papers' note
The researchers first fine-tuned computer-use models on 6,044 curated gameplay trajectories, then trained them with reinforcement learning in sixteen synthetic 2D and 3D environments. The framework adds relative pointer motion, sustained inputs and combined keyboard-and-mouse actions. The larger Player-1-35B-A3B gained 17.7 percentage points on MineDojo-300, and both models improved over supervised training on ViZDoom-10 and GameWorld. Broader coverage of synthetic environments improved transfer beyond the reinforcement-learning training suite.
HF Daily Papers' note
score 5