Gamer: A Generalist Vision-Language-Action Foundation Model for Real-Time Games
Researchers introduced Gamer, a generalist Vision-Language-Action foundation model designed to master long-horizon interactive tasks in both 2D and 3D environments. It integrates perception, memory management, and multi-step reasoning using sparse-thinking instruction tuning and reinforcement learning. This framework represents a significant step toward generalist AI agents capable of operating across complex digital and physical worlds with human-like skill. By scaling keyboard and mouse trajectories, it opens new paths for training robust agents that handle prolonged decision-making sequences. Gamer utilizes a cloud-edge inference engine to support unbounded-context execution and low online inference latency across concurrent simulated environments. It was pre-trained on tens of thousands of hours of device-level interaction data and trained with a multi-turn reinforcement learning method to mitigate accumulated rollout errors.
## BACKGROUND
Vision-Language-Action (VLA) models are multimodal foundation models that combine visual perception, language understanding, and action execution to control digital or physical agents. Long-horizon tasks refer to complex goals that require an agent to execute dozens or hundreds of sequential steps and maintain coherent intent before reaching an outcome.