Game2Player Turns 568 Web Games Into Verifiable Training Grounds for Visual Agents
UIUC-CS · hf · 2026-11-19
Researchers from UIUC, UC Santa Barbara, Georgia Tech and others introduce Game2Player, a framework that co-develops game environments and player agents to collect verifiable gameplay experience at scale.
- A Builder agent adapts existing web games, exposes internal state, and defines tasks with state-based progress measures and success checks; Coding and VLM Players attempt tasks and report issues back for repair, so play doubles as environment validation and training data collection.
- The pipeline produced 568 verified web games with 3,436 tasks, 15.7K trajectories (307 hours of gameplay), and a 50-game held-out benchmark called PlayerBench.
- Training VLM agents with progress-weighted SFT, Game2Player-9B improves progress from 14.3% to 31.2% on PlayerBench and from 30.0% to 53.1% on the external GameWorld benchmark.
More from Embodied
- Musk predicts Tesla Optimus will wash, dry and put away dishes in 12-18 months — Polymarket · 2026-11-19
- Nimo Robotics launches N1 glove and D1 hand: highest-DoF commercial UMI system for industrial robot skills — Scobleizer · 2026-11-19
- Robot Model Dream-1 Learns to Drive a Real Tesla From 1.5h of Demos — scott_e_reed · 2026-11-19
- Stack-chan desktop robot gets a screwless multi-leg armor mod, files open-sourced — alexcovo_eth · 2026-11-19
- Inceptio tops 1.2B km of driverless-truck operations, makes the case for freight-native physical AI — 量子位 · 2026-11-19
- Antigravity Robotics plugin lets anyone build MuJoCo sim tasks with prompts — DynamicWebPaige · 2026-11-19