🧑‍🔬 World Model Scientist

🧑‍🔬

World Model Scientist

Build the model: Representation → Dynamics → Prediction → Planning → Evaluation.

1. 1
   a01What is a World Model?45 min
   The big picture: what world models are, where they came from, and why they matter now.
2. 2
   a02Observation, State and POMDP60 min
   The first-principles distinction between what you see and what is true.
3. 3
   a03State Space Models75 min
   LGSSM, Kalman filtering and belief-state inference — the classical baseline.
4. 4
   a04Representation Learning60 min
   What makes a representation good enough to predict in.
5. 5
   a05Latent Dynamics75 min
   Learn a compact latent state and predict its evolution instead of pixels.
6. 6
   a06RSSM90 min
   The Recurrent State Space Model — the deterministic + stochastic backbone of Dreamer.
7. 7
   a07World Models 201845 min
   Ha & Schmidhuber’s origin paper: VAE + MDN-RNN + controller, training in dreams.
8. 8
   a08Dreamer90 min
   Actor-critic in imagination: DreamerV1/V2/V3 and the engineering of MBRL at scale.
9. 9
   a09Planning with World Models75 min
   MPC, CEM, MPPI and latent-space planning — the model becomes the environment.
10. 10
    a10Video World Models75 min
    Diffusion/flow-based generation, interactive video worlds, and closed-loop drift.
11. 11
    a10bInteractive World Models60 min
    From video generation to playable worlds: GameNGen, DIAMOND, Genie — controllability, real time and consistency.
12. 12
    a11World Model + Policy75 min
    Closing the loop: policy and value learning inside a learned model (TD-MPC, Dreamer).
13. 13
    a12OOD, Drift and Evaluation60 min
    When world models lie: out-of-distribution failure, drift, calibration and eval protocols.
14. 14
    a13Capstone: Build Your Own World Modelproject
    End-to-end project: define state / transition / action / evaluation, then build it.

░░░░░░░░░░░░░░░░░░░░ 0%0/14 completed

Next up: a01 · What is a World Model?

Why this order

The track follows the real research pipeline, not a textbook table of contents:

1. Foundations of state (a01–a03) — what a world model even is, the observation/state distinction, and the classical state-space baseline you will beat and borrow from.
2. Representation & dynamics (a04–a08) — learn a latent space worth predicting in, then the RSSM lineage: World Models 2018 → PlaNet → Dreamer.
3. Using the model (a09–a11, including a10b) — planning (MPC/CEM/MPPI), video-scale prediction, interactive world models you can play in real time, and closing the loop with a policy.
4. Trusting the model (a12) — OOD, drift and evaluation, because a world model you can't audit is a liability.
5. Capstone (a13) — build your own, with state / transition / action / evaluation made explicit.

Exit criteria

You are done with this track when you can reproduce a Dreamer / TD-MPC-class system on a new environment and write a standardized evaluation report for it (Lab 10 protocol). Career narrative: World Model Researcher / Research Engineer.

Not your track? The parallel Spatial & Embodied Engineer track builds systems instead of models — same foundations, different exit.