Gamer technical report: 10K hours of gameplay, CPT+SFT+TA-GRPO RL, Streaming-vLLM
minchoi · x · 2026-11-17
Gamer's project page details its training pipeline:
- Architecture: a pretrained multimodal model taking 5 Hz screenshots and outputting mouse displacement plus six consecutive 33 ms key-state blocks in a unified keyboard-and-mouse format.
- Three-stage training: continued pre-training (CPT) on 10K+ hours of aligned gameplay; instruction tuning (SFT) at operation/subgoal/task levels with sparse reasoning and persistent memory; online RL via TA-GRPO with environment-verified rewards on memory-linked trajectories.
- Real-time inference: Streaming-vLLM engine with persistent KV caching and early action dispatch for low-latency parallel rollouts and deployment.
- Results: completes Honkai: Star Rail's main story and solves five puzzle types with one action space.
- Open-sourced: project page, GitHub repo, Hugging Face models, technical report.
Related event: miHoYo Open-Sources Gamer, a Generalist VLA Game Agent(7 posts)→
More from Embodied
- miHoYo unveils Gamer, a generalist AI agent that plays games by watching the screen — minchoi · 2026-11-17
- Korean J&J Tech builds pharmacy humanoid on Unitree R1 for shelf checks — CyberRobooo · 2026-11-17
- Maker Builds a Vacuum Pickup Screwdriver as a Tool for a Robotic Agent — ihorbeaver · 2026-11-17
- Claude designed three working devices; the 25-degree board tilt detail stands out — nukebuddy · 2026-11-17
- Vision Pro app Passage to drop subscription for one-time purchase on Dec 1 — wattmaller1 · 2026-11-17
- First Cybercab ride in Austin: reviewer says it beats Waymo handily — Scobleizer · 2026-11-17