{
  "version": 1,
  "language": "en",
  "title": "Predict a subtask and then generate a chunk",
  "description": "The VLA selects a language subtask. That subtask and the same current context condition motor generation. The shown inference uses fixed weights.",
  "caption": "The author made this π0.5 inference diagram from Physical Intelligence’s source figure. A language-level subtask explicitly conditions continuous action generation with the current observation and original instruction. The two context cards show reuse, not two sensors. Token-level details and official illustrative images are not shown. The playback is an example, and animation time is not measured runtime.",
  "alt": "π0.5 flow: predict a language subtask, then condition a continuous action chunk on it, the current observation and original instruction.",
  "sources": [
    "https://www.pi.website/blog/pi05"
  ],
  "limitations": [
    "The playback is an example. Playback time is not measured runtime.",
    "No training or weight update occurs in this shown inference path."
  ],
  "targets": {
    "high-context": {
      "id": "mechanism-high-context",
      "kind": "node",
      "label": "Current context"
    },
    "vlm": {
      "id": "mechanism-vlm",
      "kind": "node",
      "label": "VLA prediction"
    },
    "subtask": {
      "id": "mechanism-subtask",
      "kind": "node",
      "label": "Language subtask"
    },
    "low-context": {
      "id": "mechanism-low-context",
      "kind": "node",
      "label": "Same context"
    },
    "expert": {
      "id": "mechanism-expert",
      "kind": "node",
      "label": "Action expert"
    },
    "chunk": {
      "id": "mechanism-chunk",
      "kind": "node",
      "label": "Action chunk"
    },
    "high-input": {
      "id": "mechanism-high-input",
      "kind": "edge",
      "label": "Read scene and instruction"
    },
    "choose-subtask": {
      "id": "mechanism-choose-subtask",
      "kind": "edge",
      "label": "Predict subtask"
    },
    "low-input": {
      "id": "mechanism-low-input",
      "kind": "edge",
      "label": "Use current context again"
    },
    "subtask-condition": {
      "id": "mechanism-subtask-condition",
      "kind": "edge",
      "label": "Condition on subtask"
    },
    "emit-actions": {
      "id": "mechanism-emit-actions",
      "kind": "edge",
      "label": "Generate continuous chunk"
    }
  },
  "stateLabels": {
    "stage": "Stage",
    "artifact": "Artifact"
  },
  "scenarios": [
    {
      "id": "inference",
      "title": "Subtask selection and then motor generation",
      "description": "The two context cards show reuse of the same scene and instruction.",
      "initial": {
        "stage": "Prepared",
        "artifact": "Scene and task instruction available"
      },
      "unselected": [],
      "steps": [
        {
          "title": "Read the scene and task instruction",
          "description": "The VLA reads the current scene and original task instruction.",
          "duration": 3000,
          "active": [
            "vlm"
          ],
          "paths": [
            "high-input"
          ],
          "state": {
            "stage": "High-level inference",
            "artifact": "Current scene and original instruction"
          }
        },
        {
          "title": "Select a language-level subtask",
          "description": "The VLA selects a language-level subtask.",
          "duration": 3000,
          "active": [
            "subtask"
          ],
          "paths": [
            "choose-subtask"
          ],
          "state": {
            "stage": "Subtask available",
            "artifact": "Pick up the pillow"
          }
        },
        {
          "title": "Read subtask and current context",
          "description": "The action expert reads the subtask with the same current scene and original instruction.",
          "duration": 3000,
          "active": [
            "expert"
          ],
          "paths": [
            "subtask-condition",
            "low-input"
          ],
          "state": {
            "stage": "Motor inputs available",
            "artifact": "Current context plus predicted subtask"
          }
        },
        {
          "title": "Generate a continuous action chunk",
          "description": "The action expert generates a continuous action chunk.",
          "duration": 3000,
          "active": [
            "chunk"
          ],
          "paths": [
            "emit-actions"
          ],
          "state": {
            "stage": "Motor output available",
            "artifact": "Continuous robot commands"
          }
        }
      ]
    }
  ]
}