Player-1: Learning to Play from Synthetic Experience
arXiv cs.AI · 15 Oct 2026, 03:57 AEST
Player-1 is a new framework designed to enhance game control learning through synthetic experience. It improves performance in various games by utilizing a diverse set of training environments and advanced action techniques.
Key facts
- Computer-use agents have demonstrated strong capabilities in interpreting screens and executing keyboard and mouse actions.
- We introduce Player-1, a post-training framework for learning game control beyond demonstrations through synthetic experience.
- We construct Player-1-Gym, sixteen synthetic 2D and 3D environments designed around control demands shared across games.
- The resulting Player-1-9B and Player-1-35B-A3B improve MineDojo-300 success over their SFT policies by 27.8 and 17.7 percentage points, respectively.
- Further analyses show that broader synthetic environment coverage improves transfer, while the use of composite actions that combine multiple keyboard and mouse inputs increases during RL.
Read aloud
Player-one is a new framework designed to enhance game control learning through synthetic experience. It improves performance in various games by utilizing a diverse set of training environments and advanced action techniques. Computer-use agents have demonstrated strong capabilities in interpreting screens and executing keyboard and mouse actions. We introduce Player-one, a post-training framework for learning game control beyond demonstrations through synthetic experience. We construct Player-one-Gym, sixteen synthetic 2D and 3D environments designed around control demands shared across games. The resulting Player-one-9B and Player-one-35B-A3B improve MineDojo-three hundred success over their SFT policies by twenty seven point eight and seventeen point seven percentage points, respectively. Further analyses show that broader synthetic environment coverage improves transfer, while the use of composite actions that combine multiple keyboard and mouse inputs increases during RL.
Sources in this cluster
- Player-1: Learning to Play from Synthetic Experience
arXiv cs.AI · 15 Oct 2026, 03:57 AEST