Interactive video worlds

BoxWorld.

Let World Engines Control
and Video Generators Render

Zike Wu1,3 Ahmed Hatem1,2 Chintan Trivedi2 Wallace Michel Pinto Lira2 Xuebin Qin2 Lele Wang1 Renjie Liao1,3,4

1 University of British Columbia 2 Huawei Canada 3 Vector Institute for AI 4 Canada CIFAR AI Chair

You define the rules.
The engine controls what happens.
The video model brings it to life.

Nurture & bloom
Act & change
Ride & explore
Pick & carry
Reimagine the world
Move & interact
Simple states.
Rich visual worlds.

01 / How it works

A small interface.
A clear division of work.

The engine owns the rules and state.
The renderer supplies appearance, lighting, and visual detail.

01 Author

World engine

A coding agent turns your interaction rules into executable logic.

# a user-defined behaviorif nurture_complete:
  flower.status = "open"

# the engine retains the state
User action State update
02 Expose

Object-centric state

Only the information needed to place, identify, and update each object.

Schematic layout boxes for a drone and a flowerdroneflower
Layout boxesAttributesDynamic status
Interface schematic
03 Render

Generative renderer

A causal video model renders each new state in the context of its history.

Generated greenhouse world with a drone and lotus
Updated state Video frame

Inside the renderer

The right state,
in the right place.

Frame-wise masked cross-attention routes an object’s state to the visual regions covered by its box. Persistent attributes describe what it is; dynamic status describes what is happening now.

Read the method
I

Ground objects

Learn spatial grounding from automatically annotated open-domain videos.

II

Learn state changes

Adapt to object transitions using gameplay videos and clean visual history.

III

Practice rollouts

Distill on self-generated histories for coherent, few-step streaming generation.

ResultsOn PairedStateBench, both branches of a counterfactual pair must show their intended outcome—and nothing happens when the condition is not met.

97.3%

Paired counterfactual success

83.9%

Complete-task success

20.7fps

Throughput · 832 × 480, 4 steps

02 / Key examples

Change the input.
Change the outcome.

Each pair runs the same rules on two different inputs, and the engine decides the outcome. The last example keeps the input and changes the world. Box states and held keys are shown with every video.

01 Input timing

Hold to bloom.

Stay near the lotus and hold D for two seconds: it opens. Release early and the charge resets—the flower stays closed.

Engine rulenearby + hold D for 2 s → bloom

02 Conditions

No key, no entry.

With the key, pressing D at the door opens it. Without the key, the same press is rejected and the door stays locked.

Engine rulehas key + D at the door → open

03 Contact

Grip only what you reach.

An aligned grip attaches the cargo, so the robot can carry it and set it down. Raise the arms too high and the grip closes on air; the cargo stays on the table.

Engine rulehands on cargo + grip → attached

04 Appearance

Same state, different worlds.

One engine replay, rendered as a copper drone in a greenhouse, a bee spirit in a stone hall and a ruby orb among ancient columns. Only the object and scene descriptions change.

Engine rulesame box states + new descriptions → new world

Explore the research

Let world engines control.
Let video generators render.

@misc{wu2026boxworldletworldengines,
  title={BoxWorld: Let World Engines Control and Video Generators Render},
  author={Zike Wu and Ahmed Hatem and Chintan Trivedi and Wallace Michel Pinto Lira and Xuebin Qin and Lele Wang and Renjie Liao},
  year={2026},
  eprint={2610.36371},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.36371},
}

Example

BoxWorld example