{
  "version": 1,
  "language": "en",
  "title": "Prediction before a decision",
  "description": "A world model reads history, current state and a candidate executable action. The model predicts states or observations and optional reward/value. A decision step uses the predictions. Only deployment adds environment execution and new observations. The prediction trace stops at the decision.",
  "caption": "A world model receives context and a candidate robot action. Its predicted results enter a planner, evaluator, filter or policy learner. The executed action returns new observations. Author-created explanatory diagram. Playback time is illustrative and is not measured runtime.",
  "alt": "Flow diagram: robot actions condition predictions for planning, evaluation, filtering or learning, and execution returns observations.",
  "sources": [
    {
      "path": "content/blog/2026-06-09_world-models-for-embodied-ai/post.md",
      "description": "Article mechanism and retained citations"
    },
    "https://liyang.page/wm-tutorial/",
    "https://arxiv.org/abs/2605.00080",
    "https://arxiv.org/abs/2607.00836"
  ],
  "limitations": [
    "This playback is an example. Playback duration is not measured runtime.",
    "The lower path is for deployed closed-loop operation. This path is not necessary for training-only models.",
    "Different methods can give different state, observation and optional reward/value outputs. These are not mandatory outputs for all world models."
  ],
  "targets": {
    "context": {
      "id": "mechanism-context",
      "kind": "node",
      "label": "Predictive context"
    },
    "candidate": {
      "id": "mechanism-candidate",
      "kind": "node",
      "label": "Candidate action"
    },
    "model": {
      "id": "mechanism-model",
      "kind": "node",
      "label": "World model"
    },
    "consequences": {
      "id": "mechanism-consequences",
      "kind": "node",
      "label": "Predicted results"
    },
    "consumer": {
      "id": "mechanism-consumer",
      "kind": "node",
      "label": "Decision step"
    },
    "runtime": {
      "id": "mechanism-runtime",
      "kind": "node",
      "label": "Execute + observe"
    },
    "contextModel": {
      "id": "mechanism-contextModel",
      "kind": "edge",
      "label": "supply context"
    },
    "actionModel": {
      "id": "mechanism-actionModel",
      "kind": "edge",
      "label": "supply candidate action"
    },
    "predict": {
      "id": "mechanism-predict",
      "kind": "edge",
      "label": "predict results"
    },
    "consume": {
      "id": "mechanism-consume",
      "kind": "edge",
      "label": "use prediction"
    },
    "execute": {
      "id": "mechanism-execute",
      "kind": "edge",
      "label": "do deployed action"
    },
    "feedback": {
      "id": "mechanism-feedback",
      "kind": "edge",
      "label": "read observations"
    }
  },
  "stateLabels": {
    "stage": "Stage",
    "artifact": "Artifact"
  },
  "scenarios": [
    {
      "id": "prediction",
      "title": "Prediction for a decision",
      "description": "This trace ends after a decision uses the prediction.",
      "initial": {
        "stage": "Ready",
        "artifact": "Inputs available"
      },
      "unselected": [
        "runtime",
        "execute",
        "feedback"
      ],
      "steps": [
        {
          "title": "Context and candidate inputs",
          "description": "The world model reads predictive context and an executable action candidate.",
          "duration": 3000,
          "active": [
            "context",
            "candidate"
          ],
          "paths": [],
          "state": {
            "stage": "Context and candidate inputs",
            "artifact": "History and an executable action candidate"
          }
        },
        {
          "title": "Action-conditioned prediction",
          "description": "The world model predicts results from the context and candidate action.",
          "duration": 3000,
          "active": [
            "model"
          ],
          "paths": [
            "contextModel",
            "actionModel"
          ],
          "state": {
            "stage": "Action-conditioned prediction",
            "artifact": "Conditional world-model computation"
          }
        },
        {
          "title": "Predicted result output",
          "description": "The world model gives predicted states or observations and optional reward/value.",
          "duration": 3000,
          "active": [
            "consequences"
          ],
          "paths": [
            "predict"
          ],
          "state": {
            "stage": "Predicted result output",
            "artifact": "State/observation and optional reward/value"
          }
        },
        {
          "title": "Prediction in a decision",
          "description": "The decision step uses the prediction for planning, evaluation, filtering or learning.",
          "duration": 3000,
          "active": [
            "consumer"
          ],
          "paths": [
            "consume"
          ],
          "state": {
            "stage": "Prediction in a decision",
            "artifact": "Planning, evaluation, filtering or learning"
          }
        }
      ]
    },
    {
      "id": "deployment",
      "title": "Deployed closed-loop operation",
      "description": "This trace adds environment execution and new observations after the decision.",
      "initial": {
        "stage": "Ready",
        "artifact": "Inputs available"
      },
      "unselected": [],
      "steps": [
        {
          "title": "Context and candidate inputs",
          "description": "The world model reads predictive context and an executable action candidate.",
          "duration": 3000,
          "active": [
            "context",
            "candidate"
          ],
          "paths": [],
          "state": {
            "stage": "Context and candidate inputs",
            "artifact": "History and an executable action candidate"
          }
        },
        {
          "title": "Action-conditioned prediction",
          "description": "The world model predicts results from the context and candidate action.",
          "duration": 3000,
          "active": [
            "model"
          ],
          "paths": [
            "contextModel",
            "actionModel"
          ],
          "state": {
            "stage": "Action-conditioned prediction",
            "artifact": "Conditional world-model computation"
          }
        },
        {
          "title": "Predicted result output",
          "description": "The world model gives predicted states or observations and optional reward/value.",
          "duration": 3000,
          "active": [
            "consequences"
          ],
          "paths": [
            "predict"
          ],
          "state": {
            "stage": "Predicted result output",
            "artifact": "State/observation and optional reward/value"
          }
        },
        {
          "title": "Prediction in a decision",
          "description": "The decision step uses the prediction for planning, evaluation, filtering or learning.",
          "duration": 3000,
          "active": [
            "consumer"
          ],
          "paths": [
            "consume"
          ],
          "state": {
            "stage": "Prediction in a decision",
            "artifact": "Planning, evaluation, filtering or learning"
          }
        },
        {
          "title": "Deployment action and observation",
          "description": "The deployed system does the selected action and reads new observations.",
          "duration": 3000,
          "active": [
            "runtime"
          ],
          "paths": [
            "execute"
          ],
          "state": {
            "stage": "Deployment action and observation",
            "artifact": "A deployed action gives new observations."
          }
        },
        {
          "title": "New predictive context",
          "description": "The new observations update the context for the next decision.",
          "duration": 3000,
          "active": [
            "context"
          ],
          "paths": [
            "feedback"
          ],
          "state": {
            "stage": "New predictive context",
            "artifact": "New observations for the next decision"
          }
        }
      ]
    }
  ]
}