README.md

metadata

title: GameScaling action viewer
emoji: 🎮
colorFrom: red
colorTo: green
sdk: static
app_file: index.html
short_description: A GameScaling reply as a mouse move and six 33 ms steps
license: apache-2.0
models:
  - THUSI-Lab/GameScaling

GameScaling — one 200 ms action chunk, visualised

A small static page for GameScaling (Scaling in Games: Continued Pre-Training for Embodied Agents in Diverse Virtual Worlds).

Every turn the model answers with exactly one chunk:

<|action_start|>X Y Z ; g1 ; g2 ; g3 ; g4 ; g5 ; g6<|action_end|>

X Y is the mouse displacement over the whole 200 ms in per-mille of the screen, Z the scroll amount, and g1..g6 are the keys held during each of six 33 ms sub-steps.

This page parses such a reply and draws it: the red arrow is the predicted mouse move, the strip below tells you what is held when. Press Play to step through the chunk.

There is no model here. The Doom reply is a recorded output shipped by the authors (examples/doom_battle1_action.txt); paste any other reply into the box to parse it. Live inference lives in the authors' ZeroGPU Space — Gradio/Docker Spaces need a PRO plan, and the 2B/9B checkpoints need a GPU, so this copy is the static, free-hardware option.

The parser and overlay geometry follow the authors' gamescaling package (Apache-2.0), reimplemented in plain JavaScript. Not affiliated with the authors.

Frame credits:

• doom_battle1.jpg: from the authors' repo.
• minetest.jpg: "Minetest screenshot 2" by Kotolegokot, CC BY-SA 3.0, via Wikimedia Commons.
• supertuxkart.jpg: "Supertuxkart-0.9-screenshot-2" by the SuperTuxKart dev team, CC BY 3.0, via Wikimedia Commons.