Worlds for

Observations, actions, and world state. Connected data for world model research.

DRAG TO ROTATEEXPLORE THE SPHERE

ABOUT

WorldFlux connects multi-agent worlds to your research through observations, actions, and recorded world state.

ONE SHARED ENVIRONMENT.
EVERY PERSPECTIVE CONNECTED.

Explore what each agent sees, what it does, and what changes. From Unreal Engine environments to synchronized episodes in Python.

Inside an episode

SHARED WORLDS.
INDIVIDUAL PERSPECTIVES.

FROM INTERACTION TO UNDERSTANDING.

A connected world.

Three parts of the same story.
One context, kept together.

A modular simulated city with mannequin agents.

BUILT IN UNREAL ENGINE

Start with
a shared environment.

SYNCHRONIZED CONTEXT

T × A

Time steps × agents.
Every perspective has its place.

Mannequin agents interacting in a shared courtyard.

MADE FOR YOUR WORKFLOW

From the world.
Into Python.

AI-generated environment concepts.

Unreal EngineRawCapturePythonPyTorchWorldFlux SDK

Worlds, agents
& possibilities.

Explore the engine

Three mannequin agents in a modular Unreal-style courtyard.

A world in motion.

A shared environment. Independent agents. Interactions that unfold together.

Explore the world

AI-generated simulation concept

An agent-level view through a shared architectural environment.

A different point of view.

Connect what each agent can observe to the world around it.

Explore perspectives

AI-generated simulation concept

Color-coded mannequin agents meeting along a courtyard path.

Every action. A new context.

Keep the relationships between observations, actions, and state intact.

Explore the possibility

AI-generated simulation concept

A shared world.
A possibility in every action.

Look closer

WORLDFLUX / MOTION STUDIES01 — 02

OBSERVATIONS · ACTIONS · WORLD STATE

One world.
Independent agents.

Agents navigate a shared environment. Connect their individual actions to the world state around them.

CONCEPT FILM / HIGGSFIELDAI-generated imagery · not a recorded episode

Motion study 01

Step inside.

Orbit the scene. Follow an agent.
Discover another point of view.

A modular urban simulation concept with mannequin agents. — Scripted Three.js illustration · not an Unreal Engine capture

INSIDE AN EPISODE

One world.
Many ways to see it.

A shared environment, individual viewpoints. Connect each agent’s observations with its actions and the recorded world state.

World perspective

01 / THE SHARED WORLD

A space shared by many agents.

The blue marker identifies Agent B and its viewing direction.

Agent perspective · B

02 / AGENT B · BLUE

See through the blue agent.

A camera at Agent B’s eye height shows the same courtyard from its position.

Rendering both views from the same scene.Illustrative Three.js render · not an Unreal capture.

Connect the scene to its data.

Observations, actions, and state — organized into an episode.

INSIDE AN EPISODE

SCHEMA VIEW

SHARED CAPTURE TICKt

Agent A

RGB

Depth

Seg.

Agent B

RGB

Depth

Seg.

Agent C

RGB

Depth

Seg.

observationstime × agent × sensor

Illustrative structure. Outputs depend on the capture configuration.

/01

Every perspective. One timeline.

Read the same moment across agents. Sensor outputs stay connected to the agent and capture tick that produced them.

RGBDepthSegmentationCalibration

THE SPACE BETWEEN AGENTS

Context matters.

Explore the relationships that make
multi-agent worlds worth studying.

Read observations, states, and actions across agents in the same time window. The surrounding context comes with them.

Keep the interaction intact.

WORLDS INMOTION.

EVERY MOMENT CHANGES
THE CONTEXT.

From interaction to episode

Every moment changes the context.

Overhead simulation concept with mint- and blue-accented faceless mannequin agents at different positions along a gray courtyard path. — 01
Context
Agents occupy a shared environment.

Conceptual test-level moment with mint- and blue-accented mannequin agents close to one another beside a geometric planter. — 02
Interaction
Their paths bring them into the same space.

Conceptual test-level moment with two faceless mannequin agents farther apart along the shared path. — 03
Change
Positions shift as movement continues.

AI-generated study of selected moments. Not a recorded or calibrated sequence.

FROM WORLD TO WORKFLOW

Your next
experiment starts here.

Select a time window. Inspect states, actions, and sensor references. Keep the context as you move into Python.

T×A

TIME STEPS × AGENTS
CONTEXT, KEPT TOGETHER.

quickstart.py

1from worldflux_sdk import open_episode, load_window
2from worldflux_sdk.torch import to_torch
3
4path = "data/my_episode"
5episode = open_episode(path, verify=True)
6
7window = load_window(
8    path,
9    window_size=min(32, episode.tick_count),
10)
11batch = to_torch(window)
12
13states = batch["state"]  # time × agent × feature
14actions = batch["executed_action"]
15valid = batch["state_valid"]

Requires a compatible local episode and PyTorch. Sensor outputs remain file references unless a decoder is supplied.

EVIDENCE IN EVERY EPISODE

Made to be
looked into.

Trace the original capture. Inspect validation results. Understand the record behind the scene.

Trace the source.

Connect captured records, content hashes, and validation results to their origin.

See what’s missing.

Keep validity masks and reasons alongside missing or invalid data.

Respect the lineage.

Check dataset splits against provenance to identify related episodes.

A WORLD, MADE FOR YOUR QUESTION

Start with what
you want to understand.

01
The question.
Share the interactions, observations, and scope you want to study.
02
The world.
Review the Unreal environment, capture setup, and asset permissions.
03
The episodes.
Record agreed scenarios and examine their validation results.
04
The possibility.
Review your data and connect it to your Python workflow.

A LITTLE MORE CONTEXT

Worlds forunderstanding.

WorldFlux connects multi-agent worlds to your research through observations, actions, and recorded world state.

A connected world.

Worlds, agents& possibilities.

A world in motion.

A different point of view.

Every action. A new context.

A shared world.A possibility in every action.

One world.Independent agents.

Step inside.

One world.Many ways to see it.

A space shared by many agents.

See through the blue agent.

Every perspective. One timeline.

Context matters.

Keep the interaction intact.

WORLDS INMOTION.

Every moment changes the context.

Context

Interaction

Change

Your nextexperiment starts here.

Made to belooked into.

Trace the source.

See what’s missing.

Respect the lineage.

Start with whatyou want to understand.

The question.

The world.

The episodes.

The possibility.

Good questions.Clear answers.

What data can an episode contain?

Can we use our own Unreal Engine environment?

Do we need Unreal Engine to read the data?

Does WorldFlux include a trained model or benchmark?

How do access and licensing work?

Worlds for

Worlds, agents
& possibilities.

A shared world.
A possibility in every action.

One world.
Independent agents.

One world.
Many ways to see it.

Your next
experiment starts here.

Made to be
looked into.

Start with what
you want to understand.

Good questions.
Clear answers.