{
  "schemaVersion": 1,
  "id": "atlas-lesson-diffusion-action-horizon",
  "title": "Predict action chunks and replan with feedback",
  "category": "learning",
  "format": "Atlas lesson",
  "description": "Separate generated action horizon, executed prefix and replanning interval.",
  "canonicalUrl": "https://semiaifoundry.com/learn/atlas/diffusion-action-horizon/",
  "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. Authored deterministic scalar integrator and command chunks. Rate and delay are stated assumptions; no diffusion inference or physical robot timing is measured. 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/diffusion-action-horizon-baff40bb4f62c303.json",
  "interactiveUrl": "https://semiaifoundry.com/ai-atlas/#diffusion-action-horizon",
  "related": [],
  "supersedes": [],
  "supersededBy": [],
  "contentHash": "380bff9f0111bb060aafc626e38604fe6289a27151733d960af0de3704d9b7d6",
  "contentLength": 6241,
  "jsonUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-diffusion-action-horizon.json",
  "markdownUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-diffusion-action-horizon.md",
  "sections": [
    {
      "id": "section-1",
      "title": "Predict action chunks and replan with feedback",
      "start": 45,
      "end": 432
    },
    {
      "id": "section-2",
      "title": "The mechanism",
      "start": 432,
      "end": 1437
    },
    {
      "id": "section-3",
      "title": "Science and assumptions",
      "start": 1437,
      "end": 1465
    },
    {
      "id": "section-4",
      "title": "Predict action chunks and replan with feedback · worked relation",
      "start": 1465,
      "end": 2585
    },
    {
      "id": "section-5",
      "title": "Engineering example",
      "start": 2585,
      "end": 3082
    },
    {
      "id": "section-6",
      "title": "Distinctions and tradeoffs",
      "start": 3082,
      "end": 3535
    },
    {
      "id": "section-7",
      "title": "Specifications and interfaces",
      "start": 3535,
      "end": 3956
    },
    {
      "id": "section-8",
      "title": "Interfaces",
      "start": 3956,
      "end": 4190
    },
    {
      "id": "section-9",
      "title": "Check your understanding",
      "start": 4190,
      "end": 4219
    },
    {
      "id": "section-10",
      "title": "What changes when only a prefix is executed before observing again?",
      "start": 4219,
      "end": 4780
    },
    {
      "id": "section-11",
      "title": "Evidence and scope",
      "start": 4780,
      "end": 5393
    },
    {
      "id": "section-12",
      "title": "Sources",
      "start": 5393,
      "end": 5669
    },
    {
      "id": "section-13",
      "title": "Continue exploring",
      "start": 5669,
      "end": 5692
    },
    {
      "id": "section-14",
      "title": "Foundations",
      "start": 5692,
      "end": 5948
    },
    {
      "id": "section-15",
      "title": "Related lessons",
      "start": 5948,
      "end": 6241
    }
  ],
  "citations": [
    {
      "title": "Explore in Atlas",
      "url": "https://semiaifoundry.com/ai-atlas/#diffusion-action-horizon"
    },
    {
      "title": "Diffusion Policy",
      "url": "https://arxiv.org/html/2303.04137v5"
    },
    {
      "title": "Published lesson data",
      "url": "https://semiaifoundry.com/ai-atlas/lesson-data/diffusion-action-horizon-baff40bb4f62c303.json"
    },
    {
      "title": "Atlas releases and licensing",
      "url": "https://semiaifoundry.com/ai-atlas/licenses.html"
    },
    {
      "title": "Train a denoiser and sample iteratively",
      "url": "https://semiaifoundry.com/learn/atlas/denoising-generative-objectives/"
    },
    {
      "title": "Predict transitions conditional on actions",
      "url": "https://semiaifoundry.com/learn/atlas/action-conditioned-world-model/"
    },
    {
      "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/"
    }
  ],
  "recordHash": "a65da0023eec60be61203a4bd89708ba091b3e108571187d26ced66da6dd67d8",
  "contentFormat": "text/markdown",
  "content": "model-lifecycle-world-action · Atlas v1.3.0\n\n# Predict action chunks and replan with feedback\n\nSeparate generated action horizon, executed prefix and replanning interval.\n\n[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#diffusion-action-horizon)Mathematical, data and behavior space; logical dimensions are not physical distance\n\nLearning objectiveSeparate generated action horizon, executed prefix and replanning interval.\n\n## The mechanism\n\nA diffusion action policy generates a sequence of future commands conditioned on observations. Receding-horizon execution applies only a prefix, then receives new observations and generates another sequence. The prediction horizon and action-execution horizon are therefore different parameters. Executing a longer prefix reduces replanning frequency but delays reactions to disturbances. Inference latency also consumes part of the control budget and must be measured or explicitly assumed. Diffusion Policy studies this structure; the scalar command sequence here is a worked control trace, not a trained denoising policy.\n\n1.  Record the observation time and generate an action chunk under a declared policy identity.\n2.  Execute only the selected prefix with an explicit command/period convention.\n3.  Observe disturbances and update the next planning state.\n4.  Compare closed-loop and open-loop outcomes while separating assumed inference delay from actual transition arithmetic.\n\n## Science and assumptions\n\n### Predict action chunks and replan with feedback · worked relation\n\n$$\nT_{prefix}=\\frac{H_a}{f},\\qquad x_{t+1}=x_t+a_t+d_t\n$$\n\nSeparate generated action horizon, executed prefix and replanning interval.\n\nSymbols and units\n\nT\\_prefix\n\nTime covered by executed action prefix\n\ns\n\nH\\_a\n\nExecuted commands before replanning\n\ncommands\n\nf\n\nCommand rate\n\ncommands per s\n\nx\\_t\n\nPosition\n\nm\n\na\\_t\n\nPosition increment command\n\nm per step\n\nd\\_t\n\nDisturbance at the declared step\n\nm\n\nFor $x_{\\mathrm{next}}=x+a$, start x=0 and an authored chunk \\[1,1,1,1\\] targeting x=4. After executing the first command, a disturbance adds +1, so observed x=2. Replanning with \\[1,1,0,0\\] and executing the next two commands reaches 4. Executing all three remaining original commands instead reaches 5. At an assumed 10 Hz command rate, a two-command prefix covers 0.2 s; an assumed 0.05 s inference delay is one quarter of that interval, not a measured policy latency.\n\nAssumptions Authored deterministic scalar integrator and command chunks. Rate and delay are stated assumptions; no diffusion inference or physical robot timing is measured.\n\n## Engineering example\n\nFor $x_{\\mathrm{next}}=x+a$, start x=0 and an authored chunk \\[1,1,1,1\\] targeting x=4. After executing the first command, a disturbance adds +1, so observed x=2. Replanning with \\[1,1,0,0\\] and executing the next two commands reaches 4. Executing all three remaining original commands instead reaches 5. At an assumed 10 Hz command rate, a two-command prefix covers 0.2 s; an assumed 0.05 s inference delay is one quarter of that interval, not a measured policy latency.\n\n## Distinctions and tradeoffs\n\nIt is a mistake to conclude: The original chunk becomes guaranteed correct.\n\nFrequent replanning improves responsiveness but spends more inference work. Long prefixes tolerate delay while increasing open-loop exposure to model and environment error.\n\nDistinguish predicting observations, choosing actions and controlling a changing environment. Separate generated action horizon, executed prefix and replanning interval.\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\nSeparate generated action horizon, executed prefix and replanning interval.\n\nAuthored deterministic scalar integrator and command chunks. Rate and delay are stated assumptions; no diffusion inference or physical robot timing is measured.\n\n### Interfaces\n\nInputs and state Input\n\nRecord the observation time and generate an action chunk under a declared policy identity.\n\nResult and check Output\n\nSeparate generated action horizon, executed prefix and replanning interval.\n\n## Check your understanding\n\n### What changes when only a prefix is executed before observing again?\n\n1.  The next commands can be replanned from the disturbed state.\n2.  The original chunk becomes guaranteed correct.\n3.  The policy’s weights must be retrained after every command.\n\nAnswer and reasoning\n\nThe next commands can be replanned from the disturbed state.\n\nThe next commands can be replanned from the disturbed state. Authored deterministic scalar integrator and command chunks. Rate and delay are stated assumptions; no diffusion inference or physical robot timing is measured.\n\n## Evidence and scope\n\nMode: worked example. Authored deterministic scalar integrator and command chunks. Rate and delay are stated assumptions; no diffusion inference or physical robot timing is measured. 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. Authored deterministic scalar integrator and command chunks. Rate and delay are stated assumptions; no diffusion inference or physical robot timing is measured.\n\n## Sources\n\n-   [Diffusion Policy](https://arxiv.org/html/2303.04137v5)\n\n[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/diffusion-action-horizon-baff40bb4f62c303.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)\n\n## Continue exploring\n\n### Foundations\n\n-   [Train a denoiser and sample iteratively](https://semiaifoundry.com/learn/atlas/denoising-generative-objectives/)\n-   [Predict transitions conditional on actions](https://semiaifoundry.com/learn/atlas/action-conditioned-world-model/)\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/)"
}