Tencent ARC's Player-1 trains game agents on synthetic experience, hitting 38.3% on MineDojo-300 vs 12.6% baseline
TencentARC · hf · 2026-10-15
Tencent ARC, with NUS and NTU, released Player-1, a post-training framework that teaches computer-use models game control beyond demonstrations via synthetic experience.
- Expanded action space: relative pointer motion, sustained inputs, and composite actions.
- Supervised init: fine-tuned on 6,044 curated gameplay trajectories pairing actions with instructions and visual reasoning.
- RL environments: Player-1-Gym, sixteen synthetic 2D/3D environments with programmatically verifiable rewards for multi-environment PPO.
- Results: Player-1-9B reaches 38.33% on MineDojo-300 (strongest baseline 12.63%), +27.8 points over SFT; the 35B-A3B variant gains 17.7 points, with transfer to ViZDoom-10 and GameWorld.
Broader synthetic environment coverage improves transfer, and composite-action usage increases during RL.
More from Embodied
- Gracia's new world model targets Real2Sim2Real and Physical AI — jonstephens85 · 2026-10-15
- Niantic Spatial built VPS on Pokémon Go imagery; a dev already prototyped it solo — jonstephens85 · 2026-10-15
- RoboParty launches RP1, a fully open-source full-stack humanoid robotics platform — GuanyaShi · 2026-10-15
- Survey: Towards Agentic Robots organizes the field into 17 categories across four domains — Chonghan Qin · 2026-10-15
- HumanoidToolBench benchmarks humanoid tool use from selection to mobile execution on real Unitree G1 — snupilab · 2026-10-15
- RegisterWAM adds learnable registers to world action models, hitting 81.6% on LIBERO-Plus and 93% on RoboTwin 2.0 — AutoLab-SJTU · 2026-10-15