Gamer is a general-purpose vision-language-action model for long-horizon tasks in complex interactive environments. It uses a unified architecture to integrate perception, memory management, and multi-step reasoning. It is trained on data comprising tens of thousands of hours of device-level keyboard and mouse trajectories. Sparse chain-of-thought instruction tuning, stable multi-turn reinforcement learning, and cloud-edge inference reduce redundant computation and support real-time interaction. Gamer demonstrates generalization capabilities and runtime efficiency comparable to those of skilled players across several 3D open-world games, and achieves competitive results on multiple 2D game benchmarks.
Source:Hugging Face 社区论文精选 · Read Original · 前沿热点
Leave a Reply