ActionRoPE encodes game actions as world-coordinate positions
Original title: ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models
Coordinate-conditioned noise scheduling cut revisit drift 66% versus the strongest of four baselines.
ActionRoPE is a parameter-free interface for planar-camera game world models such as top-down and isometric games. It turns action-driven camera displacement into world-coordinate rotary positions so the same location keeps one positional address over time. Coordinate-Conditioned Noise Scheduling denoises already observed regions earlier than newly revealed areas. The interface transferred across video backbones from 1.3B to 33B parameters and to out-of-distribution scenes.
Unconfirmed: The 66% drift reduction is the authors' comparison, not an independent result.
- Feature conditioning leaves motion to be learned from training data.
- Tokens for one world location share a positional address across time.
- CNS assigns token-wise schedules by known versus new regions.
- Authors compared four representative action-conditioning mechanisms.
- Transfer covered backbones from 1.3B to 33B parameters.
From the original article
1 min read at huggingface.co
Game world models have made rapid progress toward real-time interactive generation. Yet action control is still predominantly implemented as feature conditioning, leaving the model to learn from training data how each action should move the visual world. This formulation is poorly matched to a broad class of planar-camera worlds, such as top-down and isometric games.
By Zhiyang Deng, Yuzhe Wu, Zijun Lin · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.