{
  "version": 1,
  "language": "en",
  "title": "Add an optional desired scene",
  "description": "The diagram compares supported visual-subgoal conditioning and a direct mode. The optional desired image is an input with current observations.",
  "caption": "In one supported π0.7 inference mode, a lightweight world model generates a subgoal proposal from current observation and task context. The VLA conditions continuous action on the current scene and desired target. A visual subgoal is not necessary for all operating modes. The author made this explanatory diagram. The playback is an example, and animation time is not measured runtime.",
  "alt": "Optional subgoal flow: a world model supplies a proposal for a desired scene. The VLA reads the proposal and the current scene. Another trace does not use that branch.",
  "sources": [
    {
      "path": "content/blog/2026-05-25_from-pi0-to-pi07-open-pi-series/post.md",
      "description": "Article mechanism and retained citations"
    },
    "https://www.pi.website/blog/pi07"
  ],
  "limitations": [
    "The playback is an example. Playback time is not measured runtime.",
    "Visual subgoals are optional prompt conditions. They are not a necessary stage before all actions.",
    "A target image is not proof of reachability, a robot path or a successful rollout."
  ],
  "targets": {
    "context": {
      "id": "mechanism-context",
      "kind": "node",
      "label": "Scene + subtask"
    },
    "model": {
      "id": "mechanism-model",
      "kind": "node",
      "label": "Optional world model"
    },
    "subgoal": {
      "id": "mechanism-subgoal",
      "kind": "node",
      "label": "Subgoal image"
    },
    "vla": {
      "id": "mechanism-vla",
      "kind": "node",
      "label": "π0.7 VLA"
    },
    "chunk": {
      "id": "mechanism-chunk",
      "kind": "node",
      "label": "Action chunk"
    },
    "contextModel": {
      "id": "mechanism-contextModel",
      "kind": "edge",
      "label": "Condition the optional goal model"
    },
    "modelSubgoal": {
      "id": "mechanism-modelSubgoal",
      "kind": "edge",
      "label": "Make desired scene proposal"
    },
    "subgoalVLA": {
      "id": "mechanism-subgoalVLA",
      "kind": "edge",
      "label": "Supply visual target"
    },
    "contextVLA": {
      "id": "mechanism-contextVLA",
      "kind": "edge",
      "label": "Keep current scene"
    },
    "vlaChunk": {
      "id": "mechanism-vlaChunk",
      "kind": "edge",
      "label": "Generate continuous chunk"
    }
  },
  "stateLabels": {
    "stage": "Stage",
    "artifact": "Artifact"
  },
  "scenarios": [
    {
      "id": "subgoal",
      "title": "Visual-subgoal mode",
      "description": "This mode uses a proposed visual subgoal.",
      "initial": {
        "stage": "Prepared",
        "artifact": "Inputs available"
      },
      "unselected": [],
      "steps": [
        {
          "title": "Read the current scene and subtask",
          "description": "The system reads the current scene and subtask.",
          "duration": 3000,
          "active": [
            "context"
          ],
          "paths": [],
          "state": {
            "stage": "Read the current scene and subtask",
            "artifact": "Current scene and task context"
          }
        },
        {
          "title": "Make visual target proposal",
          "description": "The world model generates a visual-target proposal.",
          "duration": 3000,
          "active": [
            "model"
          ],
          "paths": [
            "contextModel"
          ],
          "state": {
            "stage": "Make visual target proposal",
            "artifact": "World-model target proposal"
          }
        },
        {
          "title": "Supply the desired subgoal image",
          "description": "The world model supplies the desired subgoal image.",
          "duration": 3000,
          "active": [
            "subgoal"
          ],
          "paths": [
            "modelSubgoal"
          ],
          "state": {
            "stage": "Supply the desired subgoal image",
            "artifact": "Desired future scene available"
          }
        },
        {
          "title": "Condition on current and desired scenes",
          "description": "The VLA conditions action on the current scene and desired target.",
          "duration": 3000,
          "active": [
            "vla"
          ],
          "paths": [
            "contextVLA",
            "subgoalVLA"
          ],
          "state": {
            "stage": "Condition on current and desired scenes",
            "artifact": "The VLA keeps current scene input with the target."
          }
        },
        {
          "title": "Generate a continuous chunk",
          "description": "The VLA generates a continuous chunk.",
          "duration": 3000,
          "active": [
            "chunk"
          ],
          "paths": [
            "vlaChunk"
          ],
          "state": {
            "stage": "Generate a continuous chunk",
            "artifact": "Conditional action chunk"
          }
        }
      ]
    },
    {
      "id": "direct",
      "title": "Direct conditioning mode",
      "description": "This mode does not use a visual subgoal.",
      "initial": {
        "stage": "Prepared",
        "artifact": "Inputs available"
      },
      "unselected": [
        "model",
        "subgoal",
        "contextModel",
        "modelSubgoal",
        "subgoalVLA"
      ],
      "steps": [
        {
          "title": "Read current conditions",
          "description": "The system reads the current scene and subtask.",
          "duration": 3000,
          "active": [
            "context"
          ],
          "paths": [],
          "state": {
            "stage": "Read current conditions",
            "artifact": "Current scene and subtask"
          }
        },
        {
          "title": "Use direct scene conditioning",
          "description": "The VLA conditions action on the current scene and subtask without a visual goal.",
          "duration": 3000,
          "active": [
            "vla"
          ],
          "paths": [
            "contextVLA"
          ],
          "state": {
            "stage": "Use direct scene conditioning",
            "artifact": "No visual-subgoal branch selected"
          }
        },
        {
          "title": "Generate a continuous chunk",
          "description": "The VLA generates a continuous chunk.",
          "duration": 3000,
          "active": [
            "chunk"
          ],
          "paths": [
            "vlaChunk"
          ],
          "state": {
            "stage": "Generate a continuous chunk",
            "artifact": "Conditional action chunk"
          }
        }
      ]
    }
  ]
}