LESSON wm.1.1 · Paper Study · 130 min

Classic World Models: VAE + RNN + Controller

Classic World Models: VAE + RNN + Controller. Learn VAE, MDN-RNN, latent dynamics, controller, and dream rollout and complete: Reproduce a compact latent-dynamics rollout. Part of

Learning objectives

  1. Explain what problem “Classic World Models: VAE + RNN + Controller” solves without hiding behind terminology.
  2. Trace the variables and causal links across VAE, MDN-RNN, latent dynamics.
  3. Complete “Reproduce a compact latent-dynamics rollout” and judge the result with evidence rather than intuition.

Core concepts

VAE

VAE sits in the classic compress-predict-control chain: a VAE encodes frames, an MDN-RNN predicts the action-conditioned next latent distribution, and a controller acts inside imagined trajectories. Real-environment return must still reveal whether the policy exploited model error.

MDN-RNN

MDN-RNN sits in the classic compress-predict-control chain: a VAE encodes frames, an MDN-RNN predicts the action-conditioned next latent distribution, and a controller acts inside imagined trajectories. Real-environment return must still reveal whether the policy exploited model error.

latent dynamics

latent dynamics sits in the classic compress-predict-control chain: a VAE encodes frames, an MDN-RNN predicts the action-conditioned next latent distribution, and a controller acts inside imagined trajectories. Real-environment return must still reveal whether the policy exploited model error.

controller

controller sits in the classic compress-predict-control chain: a VAE encodes frames, an MDN-RNN predicts the action-conditioned next latent distribution, and a controller acts inside imagined trajectories. Real-environment return must still reveal whether the policy exploited model error.

and dream rollout

and dream rollout sits in the classic compress-predict-control chain: a VAE encodes frames, an MDN-RNN predicts the action-conditioned next latent distribution, and a controller acts inside imagined trajectories. Real-environment return must still reveal whether the policy exploited model error.

Build and verify

Reproduce a compact latent-dynamics rollout

  • Predict: write the expected output, trend, or failure before running code.
  • Build: implement only the minimum components needed to answer the question.
  • Verify: compare with a baseline or trusted implementation; save seeds, parameters, and raw outputs.
  • Transfer: change one shape, dataset, scale, or workload condition and explain whether the conclusion still holds.

Open the complete interactive lesson