ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models

ActionRoPE. Action as coordinates for spatially consistent control in planar-camera game world models. By Zhiyang Deng and three others, from Tencent, the National University of Singapore, and one other institution. Published on the nineteenth of October. Its code repository is public and real: about thirty source files.

Game world models have moved quickly toward real-time interactive generation. Action control, though, is still mostly implemented as feature conditioning. The action is handed to the model as one more input, and the model is left to learn from training data how each action should move the visual world.

For a broad class of games that is a poor match. In planar-camera worlds, such as top-down and isometric games, a player's motion produces a planar camera displacement that is geometrically well defined. The geometry is already known, yet the model is asked to rediscover it from examples. Press a direction, and the view should shift by a predictable amount. Under feature conditioning, that predictable shift is something the network has to pick up statistically, from however many examples the data happens to hold.

The authors' observation is that spatially consistent control in these worlds comes from representing each action's effect directly in world coordinates. ActionRoPE is the interface that does this. It has no parameters. It converts the camera displacement driven by an action into rotary positions expressed in world coordinates. The consequence is that tokens belonging to the same place in the world keep the same positional address over time, however the camera has moved in between.

Once positions live in world coordinates, a spatial asymmetry in generation becomes visible. Regions the initial observation covered are geometrically known. Regions that are newly revealed still have to be synthesised. Treating both kinds of region the same way would ignore information the coordinates have just made explicit.

Coordinate-Conditioned Noise Scheduling, or CNS, puts that asymmetry to work. It assigns each token its own denoising schedule according to whether it is known or new. Observed regions are allowed to stabilise earlier, and new content is then generated with reference to them. What the model already saw settles first, and what it is inventing leans on that.

The comparison is against four representative action-conditioning mechanisms. ActionRoPE follows the commanded motion more faithfully, and it reduces revisit drift by sixty-six percent relative to the strongest baseline. The same interface transfers across video backbones, from one point three billion parameters up to thirty-three billion parameters, and to scenes outside the training distribution. Revisit drift is the telling measure here: it is what happens when the camera leaves a place and returns, and the world-coordinate address is exactly what should keep that place recognisable.

Instead of a learned conditioning signal, the action becomes an address. When moving the camera means shifting coordinates, a place in the world keeps its identity, and coming back to it lands on the same spot. The interface adds no parameters of its own, and it carries across backbones of very different sizes.