Courses
Course catalog

Embodied AI, Derived: from Imitation to World Action Models

VLAs act, world models imagine, and world action models do both. Derive the math (ELBO, DDPM, Flow Matching), compare the systems (RT-2, π0, GR00T N, Genie 3, V-JEPA 2, Cosmos, GAIA-2), and close-read DreamZero and Dream-1.

  1. 00

    The landscape

    1. The embodied AI map: VLAs, world models, WAMsPlanned
    2. How to study this fieldPlanned
  2. 01

    Imitation learning

    Policy learning
    1. Behavior cloning and compounding errorPlanned
    2. DAgger: fixing distribution shiftPlanned
    3. ACT: action chunking with a CVAEPlanned
    4. Diffusion Policy: multimodal actionsPlanned
  3. 02

    The generative math toolbox

    1. VAE and the ELBOPlanned
    2. DDPM from the ELBOPlanned
    3. The score view of diffusionPlanned
    4. Flow matching and rectified flowPlanned
    5. Latent video diffusion and few-step samplingPlanned
  4. 03

    Vision-language-action models

    VLA: acts
    1. RT-2: actions as tokensPlanned
    2. OpenVLA and Octo: open generalist policiesPlanned
    3. π0 and π0.5: flow-matching action expertsPlanned
    4. GR00T N: dual-system humanoid modelsPlanned
    5. China's VLAs: RDT-1B, GO-1 and morePlanned
    6. The VLA comparison chartPlanned
  5. 04

    World models: four routes

    World model: imagines
    1. Four routes to a world modelPlanned
    2. Genie 3: interactive pixel worldsPlanned
    3. V-JEPA 2: predicting in latent spacePlanned
    4. Cosmos: world foundation models for physical AIPlanned
    5. GAIA-2: controllable driving worldsPlanned
  6. 05

    World action models

    WAM: both
    1. What is a world action model?Draft
    2. DreamZero: close reading and labPlanned
    3. Dream-1: from paper to productPlanned
    4. Open problemsPlanned

Join the waitlist

Get new lessons, cards and the full roadmap poster when they ship. No spam.

Sign-ups open soon. Until then, write to contact@dpintelli.com.