Skip to main content

Core Concepts

Understanding WorldFlux architecture and key components.

Reference / Deprecated onboarding page

This page is retained for backward compatibility and deep background reading. For the supported onboarding path, use Installation and Quick Start.

What Is a World Model?​

A world model is a neural network that learns to predict how an environment evolves. It enables:

  • Planning: Predict outcomes without direct environment interaction
  • Imagination: Generate synthetic experience for training
  • Efficient Learning: Learn from fewer real interactions

Architecture Overview​

Key Components​

Encoder​

Compresses high-dimensional observations into compact latent representations.

state = model.encode(obs)  # [B, *obs_shape] -> State

State​

Core representation that captures environment state:

# DreamerV3
state.tensors["deter"]
state.tensors["stoch"]

# TD-MPC2
state.tensors["latent"]

Dynamics Model​

Predicts next latent state from current state and action:

next_state = model.transition(state, action)
next_state = model.update(state, action, obs)

Decoder​

Reconstructs observations and predicts rewards from latent states:

output = model.decode(state)
preds = output.preds

Imagination Rollouts​

trajectory = model.rollout(initial_state, actions)
# trajectory.states
# trajectory.rewards
# trajectory.continues

Training Loop​

World models learn from trajectories:

  1. Collect trajectories (obs, actions, rewards, dones)
  2. Store in ReplayBuffer
  3. Sample batches
  4. Compute losses
  5. Update model
  6. Repeat

Next Steps​