{
  "schemaVersion": 1,
  "id": "atlas-lesson-action-conditioned-world-model",
  "title": "Predict transitions conditional on actions",
  "category": "learning",
  "format": "Atlas lesson",
  "description": "Change an action while holding state fixed and compare predicted transitions with a held-out environment.",
  "canonicalUrl": "https://semiaifoundry.com/learn/atlas/action-conditioned-world-model/",
  "publishedDate": null,
  "revisionDate": "2026-10-07",
  "author": null,
  "topic": "model-lifecycle-world-action",
  "series": null,
  "license": null,
  "rightsUrl": "https://semiaifoundry.com/ai-atlas/licenses.html",
  "version": "1.3.0",
  "status": "Published educational lesson",
  "evidenceScope": "Mode: worked example. Example gain is 2 m per normalized command per step; model gain is 1.8. No clipping, noise or hidden state. This is worked arithmetic, not a trained world-model experience. Numerical examples are authored calculations, except explicitly identified automated core fixtures. Primary sources support the scoped mechanism; no human specialist review or learner outcome is implied.",
  "contentCoverage": "published-lesson",
  "artifactUrl": "https://semiaifoundry.com/ai-atlas/lesson-data/action-conditioned-world-model-edfade0d879a1d13.json",
  "interactiveUrl": "https://semiaifoundry.com/ai-atlas/#action-conditioned-world-model",
  "related": [],
  "supersedes": [],
  "supersededBy": [],
  "contentHash": "0e12dacbd86a30943e31ead249a0725a9796ebcbbcec5785f444ab7bb5bd46bc",
  "contentLength": 6626,
  "jsonUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-action-conditioned-world-model.json",
  "markdownUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-action-conditioned-world-model.md",
  "sections": [
    {
      "id": "section-1",
      "title": "Predict transitions conditional on actions",
      "start": 45,
      "end": 494
    },
    {
      "id": "section-2",
      "title": "The mechanism",
      "start": 494,
      "end": 1498
    },
    {
      "id": "section-3",
      "title": "Science and assumptions",
      "start": 1498,
      "end": 1526
    },
    {
      "id": "section-4",
      "title": "Predict transitions conditional on actions · worked relation",
      "start": 1526,
      "end": 2649
    },
    {
      "id": "section-5",
      "title": "Engineering example",
      "start": 2649,
      "end": 3076
    },
    {
      "id": "section-6",
      "title": "Distinctions and tradeoffs",
      "start": 3076,
      "end": 3539
    },
    {
      "id": "section-7",
      "title": "Specifications and interfaces",
      "start": 3539,
      "end": 4002
    },
    {
      "id": "section-8",
      "title": "Interfaces",
      "start": 4002,
      "end": 4258
    },
    {
      "id": "section-9",
      "title": "Check your understanding",
      "start": 4258,
      "end": 4287
    },
    {
      "id": "section-10",
      "title": "What distinguishes action-conditioned prediction from an action-free continuation?",
      "start": 4287,
      "end": 4868
    },
    {
      "id": "section-11",
      "title": "Evidence and scope",
      "start": 4868,
      "end": 5505
    },
    {
      "id": "section-12",
      "title": "Sources",
      "start": 5505,
      "end": 5869
    },
    {
      "id": "section-13",
      "title": "Continue exploring",
      "start": 5869,
      "end": 5892
    },
    {
      "id": "section-14",
      "title": "Foundations",
      "start": 5892,
      "end": 6215
    },
    {
      "id": "section-15",
      "title": "Related lessons",
      "start": 6215,
      "end": 6626
    }
  ],
  "citations": [
    {
      "title": "Explore in Atlas",
      "url": "https://semiaifoundry.com/ai-atlas/#action-conditioned-world-model"
    },
    {
      "title": "V-JEPA 2",
      "url": "https://arxiv.org/html/2506.09985v1"
    },
    {
      "title": "Mastering Diverse Domains through World Models",
      "url": "https://arxiv.org/html/2301.04104v2"
    },
    {
      "title": "Published lesson data",
      "url": "https://semiaifoundry.com/ai-atlas/lesson-data/action-conditioned-world-model-edfade0d879a1d13.json"
    },
    {
      "title": "Atlas releases and licensing",
      "url": "https://semiaifoundry.com/ai-atlas/licenses.html"
    },
    {
      "title": "Predict temporal representations without generating pixels",
      "url": "https://semiaifoundry.com/learn/atlas/video-latent-prediction/"
    },
    {
      "title": "Loss, gradients & optimizer",
      "url": "https://semiaifoundry.com/learn/atlas/loss-optimizer/"
    },
    {
      "title": "Evaluation & the AGI frontier",
      "url": "https://semiaifoundry.com/learn/atlas/evaluation/"
    },
    {
      "title": "Learn behavior from a model and test in the environment",
      "url": "https://semiaifoundry.com/learn/atlas/imagined-rollouts-policy-learning/"
    },
    {
      "title": "Translate learned representations into robot action units",
      "url": "https://semiaifoundry.com/learn/atlas/vision-language-action-interface/"
    },
    {
      "title": "Predict action chunks and replan with feedback",
      "url": "https://semiaifoundry.com/learn/atlas/diffusion-action-horizon/"
    }
  ],
  "recordHash": "b177ff123e547575905dfcb605972815807133d72d549582b26e88596763e549",
  "contentFormat": "text/markdown",
  "content": "model-lifecycle-world-action · Atlas v1.3.0\n\n# Predict transitions conditional on actions\n\nChange an action while holding state fixed and compare predicted transitions with a held-out environment.\n\n[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#action-conditioned-world-model)Mathematical, data and behavior space; logical dimensions are not physical distance\n\nLearning objectiveChange an action while holding state fixed and compare predicted transitions with a held-out environment.\n\n## The mechanism\n\nAn action-conditioned world model predicts consequences from state and action inputs. A visually plausible continuation from observational video is not sufficient evidence that the model distinguishes interventions. Training requires action-linked transitions with a stated coordinate system and observation timing. One-step prediction error can be small while errors compound during repeated predicted steps, especially when a policy seeks states absent from training. V-JEPA 2-AC and Dreamer use different learned-state mechanisms; the worked scalar transition below isolates their shared action-dependence requirement without replicating either.\n\n1.  Retain state, executed action and next observation as one time-aligned transition.\n2.  Fit or inspect a predictor that receives the action explicitly.\n3.  Hold initial state fixed and compare alternative action predictions.\n4.  Evaluate one-step and multi-step errors separately on held-out trajectories and policy-induced states.\n\n## Science and assumptions\n\n### Predict transitions conditional on actions · worked relation\n\n$$\n\\hat x_{t+1}=f_\\theta(\\hat x_t,a_t),\\qquad e_H=|\\hat x_{t+H}-x_{t+H}|\n$$\n\nChange an action while holding state fixed and compare predicted transitions with a held-out environment.\n\nSymbols and units\n\nx,hat x\n\nObserved and predicted scalar position in the example\n\nm\n\na\n\nDeclared normalized action command\n\ndimensionless\n\nf\\_theta\n\nLearned or supplied transition predictor\n\nfunction\n\nt,H\n\nTransition index and rollout horizon\n\nsteps\n\ne\\_H\n\nH-step position error\n\nm\n\nIn an authored environment $x_{\\mathrm{next}}=x+2a$, x=1 with action +0.5 yields 2 and action −0.5 yields 0. A model $x_{\\mathrm{hat,next}}=x+1.8a$ predicts 1.9 and 0.1, giving absolute error 0.1 each. Repeating action +0.5 for five predicted steps from x=1 gives model state 5.5 versus true state 6: error grows to 0.5. This exact linear example demonstrates accumulation, not a universal growth law.\n\nAssumptions Example gain is 2 m per normalized command per step; model gain is 1.8. No clipping, noise or hidden state. This is worked arithmetic, not a trained world-model experience.\n\n## Engineering example\n\nIn an authored environment $x_{\\mathrm{next}}=x+2a$, x=1 with action +0.5 yields 2 and action −0.5 yields 0. A model $x_{\\mathrm{hat,next}}=x+1.8a$ predicts 1.9 and 0.1, giving absolute error 0.1 each. Repeating action +0.5 for five predicted steps from x=1 gives model state 5.5 versus true state 6: error grows to 0.5. This exact linear example demonstrates accumulation, not a universal growth law.\n\n## Distinctions and tradeoffs\n\nIt is a mistake to conclude: The generated scene merely looks realistic.\n\nAction data supports control-relevant prediction but is limited by collected interventions and embodiment. A policy can exploit errors beyond that support.\n\nDistinguish predicting observations, choosing actions and controlling a changing environment. Change an action while holding state fixed and compare predicted transitions with a held-out environment.\n\n## Specifications and interfaces\n\nLearning mode\n\nworked example\n\nThe executable decoder has 423 parameters; worked extensions do not imply an additional simulator.\n\nWorked contract\n\nChange an action while holding state fixed and compare predicted transitions with a held-out environment.\n\nExample gain is 2 m per normalized command per step; model gain is 1.8. No clipping, noise or hidden state. This is worked arithmetic, not a trained world-model experience.\n\n### Interfaces\n\nInputs and state Input\n\nRetain state, executed action and next observation as one time-aligned transition.\n\nResult and check Output\n\nChange an action while holding state fixed and compare predicted transitions with a held-out environment.\n\n## Check your understanding\n\n### What distinguishes action-conditioned prediction from an action-free continuation?\n\n1.  Changing only the action can change the predicted next state.\n2.  The generated scene merely looks realistic.\n3.  Every future state is already included in the input.\n\nAnswer and reasoning\n\nChanging only the action can change the predicted next state.\n\nChanging only the action can change the predicted next state. Example gain is 2 m per normalized command per step; model gain is 1.8. No clipping, noise or hidden state. This is worked arithmetic, not a trained world-model experience.\n\n## Evidence and scope\n\nMode: worked example. Example gain is 2 m per normalized command per step; model gain is 1.8. No clipping, noise or hidden state. This is worked arithmetic, not a trained world-model experience. Numerical examples are authored calculations, except explicitly identified automated core fixtures. Primary sources support the scoped mechanism; no human specialist review or learner outcome is implied.\n\nModel lifecycle studio: worked example. Example gain is 2 m per normalized command per step; model gain is 1.8. No clipping, noise or hidden state. This is worked arithmetic, not a trained world-model experience.\n\n## Sources\n\n-   [V-JEPA 2](https://arxiv.org/html/2506.09985v1)\n-   [Mastering Diverse Domains through World Models](https://arxiv.org/html/2301.04104v2)\n\n[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/action-conditioned-world-model-edfade0d879a1d13.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)\n\n## Continue exploring\n\n### Foundations\n\n-   [Predict temporal representations without generating pixels](https://semiaifoundry.com/learn/atlas/video-latent-prediction/)\n-   [Loss, gradients & optimizer](https://semiaifoundry.com/learn/atlas/loss-optimizer/)\n-   [Evaluation & the AGI frontier](https://semiaifoundry.com/learn/atlas/evaluation/)\n\n### Related lessons\n\n-   [Learn behavior from a model and test in the environment](https://semiaifoundry.com/learn/atlas/imagined-rollouts-policy-learning/)\n-   [Translate learned representations into robot action units](https://semiaifoundry.com/learn/atlas/vision-language-action-interface/)\n-   [Predict action chunks and replan with feedback](https://semiaifoundry.com/learn/atlas/diffusion-action-horizon/)"
}