Player-1
A game control framework built on computer-use models, post-trained with reinforcement learning over synthetic environments and verifiable rewards. It supports fine-grained continuous control operations.
Timeline (last 90 days)
Player-1 is a game control framework that goes beyond learning from demonstration and performs post-training through synthetic experience.
Player-1 extends the computer-use model with three types of actions: relative pointer movement, sustained input, and composite actions.
The research team compiled 6,044 curated game trajectories for supervised initialization fine-tuning of the computer-use model.
On ViZDoom-10 and GameWorld, the reinforcement-learned models outperform the pure SFT versions.
The broader the coverage of synthetic environments, the better the model's transfer performance in new environments.
During reinforcement learning training, the model increasingly uses composite actions that simultaneously combine multiple keyboard and mouse inputs.
The synthetic experience route bypasses the bottleneck of human demonstration data, because programmatically generated training environments and verifiable rewards can be scaled on demand, but the gap between synthetic environments and real games still exists.
All Facts (7)
Player-1 is a game control framework that goes beyond learning from demonstration and performs post-training through synthetic experience.
50%UnverifiedPlayer-1 extends the computer-use model with three types of actions: relative pointer movement, sustained input, and composite actions.
50%UnverifiedThe research team compiled 6,044 curated game trajectories for supervised initialization fine-tuning of the computer-use model.
50%UnverifiedOn ViZDoom-10 and GameWorld, the reinforcement-learned models outperform the pure SFT versions.
50%UnverifiedThe broader the coverage of synthetic environments, the better the model's transfer performance in new environments.
50%UnverifiedDuring reinforcement learning training, the model increasingly uses composite actions that simultaneously combine multiple keyboard and mouse inputs.
50%UnverifiedThe synthetic experience route bypasses the bottleneck of human demonstration data, because programmatically generated training environments and verifiable rewards can be scaled on demand, but the gap between synthetic environments and real games still exists.
50%