{
  "schema_version": 3,
  "title": "Language\u2013visual handoff",
  "display_nodes": [
    {
      "id": "main",
      "label": "Main agent"
    },
    {
      "id": "subagent",
      "label": "Subagent"
    },
    {
      "id": "backend",
      "label": "Backend"
    }
  ],
  "caption_policy": "Displayed instructions and reports are concise paraphrases of recorded language handoffs. Exact original text and event references remain in each phase language_handoff. Backend-generated inspection prompts are distinguished from Main-authored instructions. Internal local roles are displayed as Subagent.",
  "visual_policy": "Native point/mask overlays, actual gripper mesh previews and measured pose-editor panels. Crops retain original pixels and link full original images. No reconstructed gripper geometry.",
  "cases": [
    {
      "id": "alin-bowl",
      "label": "RoboLab \u00b7 Bowl stacking",
      "task": "Stack the left bowl on the right bowl.",
      "episode_id": "alin-x2-20260923-bowlstack-job258573",
      "recording_date": "2026-09-23",
      "native_verifier": {
        "evaluated": true,
        "task_success": true,
        "first_grasp_success": true
      },
      "source_video_file": "alin-bowl-source.mp4",
      "source_video_sha256": "b64e75dd61c52ff99e75e0ced1df81ae4d807e1794caf0cb01e51748d11fd346",
      "source_video_fps": 7.5,
      "source_video_duration_s": 17.733333,
      "video_file": "alin-bowl-motion.mp4",
      "demo_file": "alin-bowl-demo.mp4",
      "poster_file": "alin-initial.png",
      "motion_label": "Recorded simulator motion",
      "timing_note": "Original motion and native tool images from the same execution. Reading pauses are added; model and planning waits are compressed. The point views are enlarged crops with the original images linked.",
      "selection_note": "Condensed successful path after fresh target reselection. Earlier stale-reference attempts, rejected placement draft, and narrow-angle generation are omitted; private raw logs preserve them.",
      "phases": [
        {
          "short": "Request",
          "title": "Main issues a subgoal",
          "detail": "Get a clear overhead view of both bowls.",
          "active_node": "main",
          "flow": [
            "main",
            "subagent"
          ],
          "handoff_label": "Main agent \u2192 Subagent \u00b7 request an overhead view",
          "messages": {
            "main": "Get a clear overhead view of both bowls.",
            "subagent": "Choose a view that shows both bowls.",
            "backend": "Return the scene and realize the requested viewing pose."
          },
          "visual": {
            "file": "alin-initial.png",
            "original_file": "alin-initial.png",
            "label": "A language goal",
            "kind": "Recorded RGB",
            "alt": "Stack the left bowl on the right bowl.",
            "note": "Move overhead for a clear view of the scene."
          },
          "evidence": {
            "event_sequences": [
              3,
              19,
              23,
              26,
              29,
              34,
              37,
              39
            ]
          },
          "source_start_s": 0,
          "source_end_s": 1.4666666666666666,
          "start_s": 0,
          "end_s": 1.5,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 19,
              "text": "overhead observation of the bowls",
              "representation": "Recorded original text",
              "tool": "delegate_waypoint"
            }
          }
        },
        {
          "short": "Point",
          "title": "Point to the left bowl",
          "detail": "Select the bowl at (418, 625) in the wrist image.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 point to the left bowl",
          "messages": {
            "main": "Find the left bowl in the current image.",
            "subagent": "Left bowl selected; point, mask and 3D geometry returned.",
            "backend": "Return the selected point, segmentation mask and measured geometry."
          },
          "visual": {
            "file": "alin-target-focus.jpg",
            "original_file": "alin-target.png",
            "label": "Point to the left bowl",
            "kind": "Native point + mask",
            "alt": "Select the bowl at (418, 625) in the wrist image.",
            "note": "The original red cross marks the selected point; cyan shows the returned segmentation. Enlarged crop; open the original for the full view."
          },
          "evidence": {
            "event_sequences": [
              79,
              83,
              84,
              86,
              88
            ]
          },
          "source_start_s": 1.4666666666666666,
          "source_end_s": 1.4666666666666666,
          "selection": {
            "view_id": "robot0_eye_in_hand",
            "normalization": "0..1000",
            "u": 418,
            "v": 625,
            "normalized_xy": [
              0.418,
              0.625
            ],
            "native_pixel_xy": [
              267,
              224
            ]
          },
          "start_s": 1.5,
          "end_s": 5.5,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 79,
              "text": "the left bowl in the current observation",
              "representation": "Recorded original text",
              "tool": "delegate_point"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 86,
              "text": "Selected the left bowl, successfully segmented and paired with 3D geometry across views.",
              "representation": "Recorded original text",
              "handoff_event_seq": 88
            }
          }
        },
        {
          "short": "Grasps",
          "title": "Compare three grasp poses",
          "detail": "Inspect three Contact-GraspNet proposals.",
          "active_node": "backend",
          "flow": [
            "backend",
            "subagent"
          ],
          "handoff_label": "Backend \u2192 Subagent \u00b7 compare three grasp poses",
          "messages": {
            "main": "Grasp the left bowl by its rim or center, then lift it for transport.",
            "subagent": "Compare three gripper candidates for a rim grasp.",
            "backend": "Return three feasible gripper poses with native cyan mesh previews."
          },
          "visual": {
            "file": "alin-grasp-1.png",
            "original_file": "alin-grasp-1.png",
            "label": "Compare three grasp poses",
            "kind": "Native mesh projections",
            "alt": "Inspect three Contact-GraspNet proposals.",
            "note": "Each cyan mesh is the actual backend preview. Grasp 1 is the recorded selection; occlusion is not tested by these overlays. Use the thumbnails below to inspect each original proposal."
          },
          "evidence": {
            "event_sequences": [
              94,
              101,
              102,
              104,
              105,
              107,
              109,
              110,
              116,
              117,
              119,
              121,
              123
            ]
          },
          "source_start_s": 1.4666666666666666,
          "source_end_s": 1.4666666666666666,
          "candidates": [
            {
              "file": "alin-grasp-1.png",
              "label": "Grasp 1 \u00b7 selected",
              "selected": true,
              "source_candidate_index": 0,
              "score": 0.23478765785694122
            },
            {
              "file": "alin-grasp-2.png",
              "label": "Grasp 2",
              "selected": false,
              "source_candidate_index": 2,
              "score": 0.16471675038337708
            },
            {
              "file": "alin-grasp-3.png",
              "label": "Grasp 3",
              "selected": false,
              "source_candidate_index": 1,
              "score": 0.1591050624847412
            }
          ],
          "start_s": 5.5,
          "end_s": 9.5,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 94,
              "text": "Grasp the left bowl rim or center with transport mode",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            }
          }
        },
        {
          "short": "Inspect",
          "title": "Inspect the chosen gripper",
          "detail": "Accept Grasp 1 after inspecting its measured geometry.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "main"
          ],
          "handoff_label": "Subagent \u2192 Main agent \u00b7 inspect the chosen gripper",
          "messages": {
            "main": "Grasp the left bowl by its rim or center, then lift it for transport.",
            "subagent": "Rim grasp inspected; clearance and lift height are suitable.",
            "backend": "Return the measured point cloud and gripper inspection views."
          },
          "visual": {
            "file": "alin-grasp-inspection.jpg",
            "original_file": "alin-grasp-editor.png",
            "label": "Inspect the chosen gripper",
            "kind": "Native point-cloud view",
            "alt": "Accept Grasp 1 after inspecting its measured geometry.",
            "note": "Original SIDE / TOP / CLOSING PLANE panels. Nominal gripper opening: 44 mm. Rotation examples in the full editor are explanatory; no rotation edit was applied."
          },
          "evidence": {
            "event_sequences": [
              94,
              101,
              102,
              104,
              105,
              107,
              109,
              110,
              116,
              117,
              119,
              121,
              123
            ]
          },
          "source_start_s": 1.4666666666666666,
          "source_end_s": 1.4666666666666666,
          "selected_candidate_index": 0,
          "nominal_opening_mm": 44,
          "start_s": 9.5,
          "end_s": 13.5,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 94,
              "text": "Grasp the left bowl rim or center with transport mode",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 119,
              "text": "Inspected candidate g_179c18d6a70d429098169b12f792284c: the grasp establishes a rim pinch on the left bowl with the contact center positioned at Z=0.0487 m, well clear of the table surface and appropriately aligned for lifting with transport post-pick height at 0.2687 m.",
              "representation": "Recorded original text",
              "handoff_event_seq": 121
            },
            "runtime_instruction": {
              "sender": "Backend",
              "recipient": "Subagent",
              "event_seq": 110,
              "text": "Grasp the left bowl rim or center with transport mode Inspect this selected pose before execution; preserve explicit transit heights. Geometric mean/median grasps remain downward, yaw-only. Assess the selected pose. Make a small useful correction, keep it if suitable, or return needs_observation with a helpful viewing direction when new geometry is preferable to a large pose change. Scene relaxation is optional and only useful if the pose remains appropriate.",
              "representation": "Recorded original text",
              "tool": "delegate_refiner"
            }
          }
        },
        {
          "short": "Approach",
          "title": "Approach with open fingers",
          "detail": "Move to the accepted gripper pose with the fingers open.",
          "active_node": "backend",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 approach with open fingers",
          "messages": {
            "main": "Grasp the left bowl by its rim or center, then lift it for transport.",
            "subagent": "Chosen rim grasp returned for execution.",
            "backend": "Move the open gripper to the selected pose."
          },
          "visual": {
            "file": "alin-grasp-1.png",
            "original_file": "alin-grasp-1.png",
            "label": "Approach with open fingers",
            "kind": "Native selected pose",
            "alt": "Move to the accepted gripper pose with the fingers open.",
            "note": "Original approach motion plays on the left; the accepted cyan pose stays visible on the right."
          },
          "evidence": {
            "event_sequences": [
              125
            ]
          },
          "source_start_s": 1.4666666666666666,
          "source_end_s": 7.733333333333333,
          "start_s": 13.5,
          "end_s": 19.777778,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 94,
              "text": "Grasp the left bowl rim or center with transport mode",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 107,
              "text": "Candidate g_179c18d6a70d429098169b12f792284c provides a viable rim pinch grasp on the left bowl with proper approach clearance, jaw placement across the bowl wall, and post-pick transit height 0.22 m above grasp Z for transport mode.",
              "representation": "Recorded original text",
              "handoff_event_seq": 109
            }
          }
        },
        {
          "short": "Check",
          "title": "Check before closing",
          "detail": "Inspect the fresh paused pregrasp; continue without a pose nudge.",
          "active_node": "subagent",
          "flow": [
            "backend",
            "subagent"
          ],
          "handoff_label": "Backend \u2192 Subagent \u00b7 check before closing",
          "messages": {
            "main": "Grasp the left bowl by its rim or center, then lift it for transport.",
            "subagent": "Fresh pregrasp checked; continue without a pose nudge.",
            "backend": "Return fresh wrist RGB and the paused gripper preview."
          },
          "visual": {
            "file": "alin-pregrasp.png",
            "original_file": "alin-pregrasp.png",
            "label": "Check before closing",
            "kind": "Native paused preview",
            "alt": "Inspect the fresh paused pregrasp; continue without a pose nudge.",
            "note": "Fresh preview of the pending grasp on the measured scene. The recorded decision is continue; no grasp edit is applied."
          },
          "evidence": {
            "event_sequences": [
              129,
              130,
              132
            ]
          },
          "source_start_s": 7.733333333333333,
          "source_end_s": 7.733333333333333,
          "candidates": [
            {
              "file": "alin-pregrasp.png",
              "label": "Paused mesh",
              "selected": true
            },
            {
              "file": "alin-pregrasp-wrist.png",
              "label": "Fresh wrist RGB",
              "selected": false
            }
          ],
          "start_s": 19.777778,
          "end_s": 23.777778,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 94,
              "text": "Grasp the left bowl rim or center with transport mode",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 130,
              "text": "The open gripper fingers cleanly straddle the left bowl rim as verified in the eye-in-hand view and inflight preview; alignment is suitable to close.",
              "representation": "Recorded original text",
              "handoff_event_seq": 132
            },
            "runtime_instruction": {
              "sender": "Backend",
              "recipient": "Subagent",
              "event_seq": 129,
              "text": "Inspect the fresh open-hand pregrasp and pending pose. Correct if useful, then explicitly continue to close or abort. Mean/median grasps allow translation and yaw only. Original delegated goal: Grasp the left bowl rim or center with transport mode",
              "representation": "Recorded original text"
            }
          }
        },
        {
          "short": "Lift",
          "title": "Close and lift",
          "detail": "Close the gripper and lift the selected bowl.",
          "active_node": "backend",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 close and lift",
          "messages": {
            "main": "Grasp the left bowl by its rim or center, then lift it for transport.",
            "subagent": "Continue with the inspected rim grasp.",
            "backend": "Close, lift, and return a fresh observation."
          },
          "visual": {
            "file": "alin-pregrasp-wrist.png",
            "original_file": "alin-pregrasp-wrist.png",
            "label": "Close and lift",
            "kind": "Fresh wrist RGB",
            "alt": "Close the gripper and lift the selected bowl.",
            "note": "The original motion continues from the inspected open-hand pose."
          },
          "evidence": {
            "event_sequences": [
              135,
              136,
              137,
              138
            ]
          },
          "source_start_s": 7.733333333333333,
          "source_end_s": 13.8,
          "start_s": 23.777778,
          "end_s": 29.833333,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 94,
              "text": "Grasp the left bowl rim or center with transport mode",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 130,
              "text": "The open gripper fingers cleanly straddle the left bowl rim as verified in the eye-in-hand view and inflight preview; alignment is suitable to close.",
              "representation": "Recorded original text",
              "handoff_event_seq": 132
            }
          }
        },
        {
          "short": "Receiver",
          "title": "Point to the receiving bowl",
          "detail": "Select the receiving bowl at (630, 617).",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 point to the receiving bowl",
          "messages": {
            "main": "Find the receiving bowl on the table.",
            "subagent": "Receiving bowl selected and measured.",
            "backend": "Return the receiving-bowl mask and destination geometry."
          },
          "visual": {
            "file": "alin-destination-focus.jpg",
            "original_file": "alin-destination.png",
            "label": "Point to the receiving bowl",
            "kind": "Native point + mask",
            "alt": "Select the receiving bowl at (630, 617).",
            "note": "Select the destination in the fresh post-grasp image. The native cross and mask are preserved."
          },
          "evidence": {
            "event_sequences": [
              139,
              143,
              144,
              146,
              148,
              149,
              150
            ]
          },
          "source_start_s": 13.8,
          "source_end_s": 13.8,
          "selection": {
            "view_id": "agentview",
            "normalization": "0..1000",
            "u": 630,
            "v": 617,
            "normalized_xy": [
              0.63,
              0.617
            ],
            "native_pixel_xy": [
              403,
              222
            ]
          },
          "start_s": 29.833333,
          "end_s": 33.833333,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 139,
              "text": "the bowl on the table",
              "representation": "Recorded original text",
              "tool": "delegate_destination"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 146,
              "text": "The destination bowl resting on the table is cleanly segmented and measured, centered on the interior of the bowl.",
              "representation": "Recorded original text",
              "handoff_event_seq": 148
            }
          }
        },
        {
          "short": "Pose",
          "title": "Inspect the placement draft",
          "detail": "Compare the held-bowl footprint with the measured receiving bowl.",
          "active_node": "subagent",
          "flow": [
            "backend",
            "subagent"
          ],
          "handoff_label": "Backend \u2192 Subagent \u00b7 inspect the placement draft",
          "messages": {
            "main": "Place the held bowl directly inside the right bowl.",
            "subagent": "Inspect the placement over the receiving bowl.",
            "backend": "Return the native placement preview."
          },
          "visual": {
            "file": "alin-place-draft.png",
            "original_file": "alin-place-draft.png",
            "label": "Inspect the placement draft",
            "kind": "Native placement preview",
            "alt": "Compare the held-bowl footprint with the measured receiving bowl.",
            "note": "The agent inspects the mean-reference draft before adjusting its translation."
          },
          "evidence": {
            "event_sequences": [
              153,
              158,
              159,
              167,
              168
            ]
          },
          "source_start_s": 13.8,
          "source_end_s": 13.8,
          "start_s": 33.833333,
          "end_s": 37.833333,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 153,
              "text": "Stack the held bowl directly into the right bowl",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            }
          }
        },
        {
          "short": "Shift",
          "title": "Shift the placement",
          "detail": "Adjust X by +20 mm and Y by \u221230 mm.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 shift the placement",
          "messages": {
            "main": "Place the held bowl directly inside the right bowl.",
            "subagent": "Shift the placement by +20 / \u221230 mm.",
            "backend": "Apply the requested translation and return the updated preview."
          },
          "visual": {
            "file": "alin-place-shift.png",
            "original_file": "alin-place-shift.png",
            "label": "Shift the placement",
            "kind": "Native edited pose",
            "alt": "Adjust X by +20 mm and Y by \u221230 mm.",
            "note": "First recorded translation adjustment. Rotation stays unchanged."
          },
          "evidence": {
            "event_sequences": [
              170,
              171
            ]
          },
          "source_start_s": 13.8,
          "source_end_s": 13.8,
          "adjustment_mm": [
            20,
            -30,
            0
          ],
          "start_s": 37.833333,
          "end_s": 41.833333,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 153,
              "text": "Stack the held bowl directly into the right bowl",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            }
          }
        },
        {
          "short": "Adjust",
          "title": "Inspect the final placement",
          "detail": "Refine X by +1 mm and Y by \u221212 mm, then accept the pose.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "main"
          ],
          "handoff_label": "Subagent \u2192 Main agent \u00b7 inspect the final placement",
          "messages": {
            "main": "Place the held bowl directly inside the right bowl.",
            "subagent": "Final placement accepted after a further +1 / \u221212 mm shift.",
            "backend": "Return the final pose preview with total shift +21 / \u221242 mm."
          },
          "visual": {
            "file": "alin-place-final.png",
            "original_file": "alin-place-editor.png",
            "label": "Inspect the final placement",
            "kind": "Native edited pose",
            "alt": "Refine X by +1 mm and Y by \u221212 mm, then accept the pose.",
            "note": "Total translation: +21 / \u221242 / 0 mm; rotation: 0\u00b0. The original editor records these totals."
          },
          "evidence": {
            "event_sequences": [
              173,
              174,
              176,
              178,
              180,
              184,
              185,
              187,
              189
            ]
          },
          "source_start_s": 13.8,
          "source_end_s": 13.8,
          "adjustment_mm": [
            1,
            -12,
            0
          ],
          "cumulative_translation_mm": [
            21,
            -42,
            0
          ],
          "start_s": 41.833333,
          "end_s": 45.833333,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 153,
              "text": "Stack the held bowl directly into the right bowl",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 187,
              "text": "The inspected placement candidate aligns the held bowl centered directly over the destination bowl (offset dx=0mm, dy=0mm) with valid clearance for stacking.",
              "representation": "Recorded original text",
              "handoff_event_seq": 189
            },
            "runtime_instruction": {
              "sender": "Backend",
              "recipient": "Subagent",
              "event_seq": 181,
              "text": "Stack the held bowl directly into the right bowl Review the pending placement before movement.",
              "representation": "Recorded original text"
            }
          }
        },
        {
          "short": "Place",
          "title": "Execute the inspected placement",
          "detail": "Move the bowl to the accepted final placement.",
          "active_node": "backend",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 execute the inspected placement",
          "messages": {
            "main": "Place the held bowl directly inside the right bowl.",
            "subagent": "Return the accepted placement for execution.",
            "backend": "Transport the bowl to the accepted final placement."
          },
          "visual": {
            "file": "alin-place-final.png",
            "original_file": "alin-place-final.png",
            "label": "Execute the inspected placement",
            "kind": "Native accepted pose",
            "alt": "Move the bowl to the accepted final placement.",
            "note": "Original transport and placement motion, from the same episode as all previews."
          },
          "evidence": {
            "event_sequences": [
              190,
              191,
              193
            ]
          },
          "source_start_s": 13.8,
          "source_end_s": 17.6,
          "start_s": 45.833333,
          "end_s": 49.611111,
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 153,
              "text": "Stack the held bowl directly into the right bowl",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 187,
              "text": "The inspected placement candidate aligns the held bowl centered directly over the destination bowl (offset dx=0mm, dy=0mm) with valid clearance for stacking.",
              "representation": "Recorded original text",
              "handoff_event_seq": 189
            }
          }
        },
        {
          "short": "Outcome",
          "title": "Native task outcome",
          "detail": "The native task verifier records success.",
          "active_node": "backend",
          "flow": [
            "backend",
            "main"
          ],
          "handoff_label": "Backend \u2192 Main agent \u00b7 native task outcome",
          "messages": {
            "main": "Place the held bowl directly inside the right bowl.",
            "subagent": "Final placement accepted after the recorded adjustments.",
            "backend": "Native task verifier: success."
          },
          "visual": {
            "file": "alin-outcome.png",
            "original_file": "alin-outcome.png",
            "label": "Native task outcome",
            "kind": "Recorded final RGB",
            "alt": "The native task verifier records success.",
            "note": "The final sensor is one physics step beyond the last video frame. The release request ends with the episode; successful release is not claimed."
          },
          "evidence": {
            "event_sequences": [
              194,
              196,
              197,
              198
            ]
          },
          "source_start_s": 17.6,
          "source_end_s": 17.6,
          "start_s": 49.611111,
          "end_s": 52.611111,
          "node_labels": {
            "main": "Last instruction \u2192 Subagent",
            "subagent": "Latest report \u2192 Main agent",
            "backend": "Task outcome"
          },
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 153,
              "text": "Stack the held bowl directly into the right bowl",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 187,
              "text": "The inspected placement candidate aligns the held bowl centered directly over the destination bowl (offset dx=0mm, dy=0mm) with valid clearance for stacking.",
              "representation": "Recorded original text",
              "handoff_event_seq": 189
            }
          }
        }
      ],
      "presentation_duration_s": 52.611111,
      "language_source_sha256": "e70a48b370e8306f98ade961e1eccfddafc001ef2663c69b85db92a166fe44cc"
    },
    {
      "id": "panda-pick-place",
      "label": "Panda \u00b7 Pick & place",
      "task": "Pick up the orange cube and place it in the tray.",
      "episode_id": "gui_20260926_112917_91cce9d7",
      "video_file": "panda-pick-place-motion.mp4",
      "demo_file": "panda-pick-place-demo.mp4",
      "poster_file": "panda-pick-capture_0001.png",
      "motion_label": "Recorded camera observations",
      "timing_note": "Camera observations paired with native tool images. Recorded points and gripper poses are shown alongside the physical scene.",
      "camera_frames": [
        {
          "file": "panda-pick-capture_0001.png",
          "capture_id": "capture_0001",
          "sha256": "fcd6d2f66c565fa6d25725b22fe10b9aeaa233b3bd8245cfedeb9d03a2b6b54c"
        },
        {
          "file": "panda-pick-capture_0002.png",
          "capture_id": "capture_0002",
          "sha256": "75d024df95379eacacd40f17f6010d804db96aa4a9b26577cefaacf3de9be3e1"
        },
        {
          "file": "panda-pick-capture_0003.png",
          "capture_id": "capture_0003",
          "sha256": "fe4c7b235036bb6c79297e8d1e67af511e92c4d43896dde810ad39d42672a1c7"
        },
        {
          "file": "panda-pick-capture_0004.png",
          "capture_id": "capture_0004",
          "sha256": "358320f773e45706d524860d475478d01cdff2fbaaca4eb5cec0fe6454322378"
        },
        {
          "file": "panda-pick-capture_0005.png",
          "capture_id": "capture_0005",
          "sha256": "db83059a3eaad85c9cc77a3a5592cba4b68728198b42feb3533edc45e0f581fa"
        },
        {
          "file": "panda-pick-capture_0006.png",
          "capture_id": "capture_0006",
          "sha256": "500baa65ec5c97380b9eb49ad1a68972c617d7efdf56ae76c19b985ecf38f05b"
        },
        {
          "file": "panda-pick-capture_0007.png",
          "capture_id": "capture_0007",
          "sha256": "a50401b908fcc3a8eeeef311ce2cbf7dd2a9e946b1bcfc59285bc0c505af1da5"
        },
        {
          "file": "panda-pick-capture_0008.png",
          "capture_id": "capture_0008",
          "sha256": "b7a4c7898e7970bbb1108d25eae40ffb9d21078a05ae21ee47b06e3a64aaa49d"
        },
        {
          "file": "panda-pick-capture_0009.png",
          "capture_id": "capture_0009",
          "sha256": "6c7f01ca951fea2acdb79c251243effc442970bbd5327ca433c13d86a47607be"
        },
        {
          "file": "panda-pick-capture_0010.png",
          "capture_id": "capture_0010",
          "sha256": "293343b51d0ffc331939d44ffdbcb740c1c065a122cdc7e832af386c9aeba82a"
        },
        {
          "file": "panda-pick-capture_0011.png",
          "capture_id": "capture_0011",
          "sha256": "8bbad2754659f90265425e79c5fc3405f3b6984215641acb9c907d5f8c28268b"
        },
        {
          "file": "panda-pick-capture_0012.png",
          "capture_id": "capture_0012",
          "sha256": "c948a11be7daff7670db518fb722c0d9fd3c8f4e1d522e3498395273daed0420"
        }
      ],
      "outcome": {
        "agent_status": "completed",
        "agent_verified": false,
        "physical_task_success_flag": false,
        "display_label": "Agent-reported completion; final camera observation shown",
        "independent_physical_success_verifier": false
      },
      "selection_note": "Compact accepted sequence; earlier proposal attempts and tool-contract retries omitted. All displayed tool images and camera samples come from the same execution.",
      "grasp_source": "mean",
      "selected_candidate": "g_009",
      "phases": [
        {
          "short": "Request",
          "title": "Main issues a subgoal",
          "detail": "Select the orange cube in the current image.",
          "active_node": "main",
          "flow": [
            "main",
            "subagent"
          ],
          "handoff_label": "Main agent \u2192 Subagent \u00b7 request the visual target",
          "messages": {
            "main": "Select the orange cube in the current image.",
            "subagent": "Point to the cube and return the measured region.",
            "backend": "Return the current robot camera views."
          },
          "visual": {
            "file": "panda-pick-phase-01.png",
            "original_file": "panda-pick-phase-01.png",
            "label": "Language goal",
            "kind": "Recorded RGB",
            "alt": "The main agent asks a subagent to locate the orange cube.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              3,
              4
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "fcd6d2f66c565fa6d25725b22fe10b9aeaa233b3bd8245cfedeb9d03a2b6b54c"
            }
          ],
          "start_s": 0,
          "end_s": 2.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 4,
              "text": "select the orange cube",
              "representation": "Recorded original text",
              "tool": "delegate_point"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Point",
          "title": "Point to measured mask",
          "detail": "The subagent chooses a point on the orange cube. The backend returns the measured SAM region.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 point to measured mask",
          "messages": {
            "main": "Select the orange cube in the current image.",
            "subagent": "Orange cube selected; mask and measured geometry returned.",
            "backend": "Return the selected point, segmentation mask and 3D geometry."
          },
          "visual": {
            "file": "panda-pick-target-focus.jpg",
            "original_file": "panda-pick-phase-02.png",
            "label": "Point to measured mask",
            "kind": "Native point + mask",
            "alt": "The subagent chooses a point on the orange cube. The backend returns the measured SAM region.",
            "note": "Original selection cross and segmentation mask, enlarged from the recorded camera view."
          },
          "evidence": {
            "event_sequences": [
              8,
              9,
              11,
              13
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "fcd6d2f66c565fa6d25725b22fe10b9aeaa233b3bd8245cfedeb9d03a2b6b54c"
            }
          ],
          "selection": {
            "normalized_uv": [
              502,
              810
            ],
            "normalized_coordinate_range": [
              0,
              1000
            ],
            "native_marker_pixel_xy": [
              642,
              582
            ],
            "depth_refinement_seed_pixel_xy": [
              644,
              582
            ],
            "observation_id": "obs_0001",
            "point_ref": "pt_002"
          },
          "start_s": 2.0,
          "end_s": 6.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 4,
              "text": "select the orange cube",
              "representation": "Recorded original text",
              "tool": "delegate_point"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 13,
              "text": "Orange cube cleanly segmented and measured.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Grasps",
          "title": "Grasp candidates",
          "detail": "The backend returns two feasible geometric mean grasp candidates after earlier proposals are rejected.",
          "active_node": "backend",
          "flow": [
            "backend",
            "subagent"
          ],
          "handoff_label": "Backend \u2192 Subagent \u00b7 grasp candidates",
          "messages": {
            "main": "Grasp the orange cube from above and lift it at least 20 cm for transport.",
            "subagent": "Compare 2 feasible geometric gripper poses.",
            "backend": "Return 2 geometric mean grasp proposals."
          },
          "visual": {
            "file": "panda-pick-g_009.png",
            "original_file": "panda-pick-g_009.png",
            "label": "Grasp candidates",
            "kind": "Native mesh projections",
            "alt": "The backend returns two feasible geometric mean grasp candidates after earlier proposals are rejected.",
            "note": "2 feasible geometric mean proposals. g_009 is the recorded selection. Preview occlusion is not tested. Use the thumbnails below to inspect each original proposal."
          },
          "evidence": {
            "event_sequences": [
              16,
              20,
              21,
              23,
              24,
              26,
              27
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "fcd6d2f66c565fa6d25725b22fe10b9aeaa233b3bd8245cfedeb9d03a2b6b54c"
            }
          ],
          "candidates": [
            {
              "file": "panda-pick-g_008.png",
              "label": "g_008",
              "selected": false,
              "yaw_deg": -45.0,
              "source": "mean",
              "contact_center_xyz_m": [
                0.41921181172132493,
                -0.09160453364253045,
                0.11
              ],
              "transit": {
                "pre_pick_z_m": 0.5467970209939896,
                "post_pick_z_m": 0.32
              }
            },
            {
              "file": "panda-pick-g_009.png",
              "label": "g_009 \u00b7 selected",
              "selected": true,
              "yaw_deg": 0.0,
              "source": "mean",
              "contact_center_xyz_m": [
                0.41921181172132493,
                -0.09160453364253045,
                0.11
              ],
              "transit": {
                "pre_pick_z_m": 0.5467970209939896,
                "post_pick_z_m": 0.32
              }
            }
          ],
          "start_s": 6.0,
          "end_s": 10.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "pick up the orange cube using a top-down grasp with transport lift height at least 0.20m above grasp",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Inspect",
          "title": "Inspect and choose",
          "detail": "The subagent inspects both feasible candidates and returns g_009, with a 0.11 m contact height and a 0.32 m lift height.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "main"
          ],
          "handoff_label": "Subagent \u2192 Main agent \u00b7 inspect and choose",
          "messages": {
            "main": "Grasp the orange cube from above and lift it at least 20 cm for transport.",
            "subagent": "Choose g_009: top-down grasp with sufficient transport lift.",
            "backend": "Return native gripper projections and point-cloud inspection views."
          },
          "visual": {
            "file": "panda-pick-grasp-inspection.jpg",
            "original_file": "panda-pick-phase-04.png",
            "label": "Inspect and choose",
            "kind": "Native point-cloud view",
            "alt": "The subagent inspects both feasible candidates and returns g_009, with a 0.11 m contact height and a 0.32 m lift height.",
            "note": "SIDE / TOP / CLOSING PLANE panels of g_009. Opening: 80 mm. Lower rotation examples are explanatory; no pose edit was applied."
          },
          "evidence": {
            "event_sequences": [
              29,
              30,
              32,
              33,
              35,
              37
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "fcd6d2f66c565fa6d25725b22fe10b9aeaa233b3bd8245cfedeb9d03a2b6b54c"
            }
          ],
          "selected_candidate": "g_009",
          "nominal_opening_mm": 80,
          "start_s": 10.0,
          "end_s": 14.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "pick up the orange cube using a top-down grasp with transport lift height at least 0.20m above grasp",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 37,
              "text": "Candidate g_009 provides a viable, top-down executable grasp aligned well with the orange cube, with contact center Z at 0.11 m and post-pick lift height at 0.32 m (0.21 m above grasp height), satisfying the transport lift requirement.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Approach",
          "title": "Approach",
          "detail": "The backend validates g_009 and starts the physical grasp primitive.",
          "active_node": "backend",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 approach",
          "messages": {
            "main": "Grasp the orange cube from above and lift it at least 20 cm for transport.",
            "subagent": "Return g_009 for validated physical execution.",
            "backend": "Validate g_009 and start the grasp primitive."
          },
          "visual": {
            "file": "panda-pick-g_009.png",
            "original_file": "panda-pick-g_009.png",
            "label": "Approach",
            "kind": "Native selected pose",
            "alt": "The backend validates g_009 and starts the physical grasp primitive.",
            "note": "The backend validates g_009 and starts the physical grasp primitive."
          },
          "evidence": {
            "event_sequences": [
              38,
              39,
              40
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0002.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0002.png",
              "capture_id": "capture_0002",
              "sha256": "75d024df95379eacacd40f17f6010d804db96aa4a9b26577cefaacf3de9be3e1"
            }
          ],
          "start_s": 14.0,
          "end_s": 17.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "pick up the orange cube using a top-down grasp with transport lift height at least 0.20m above grasp",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 37,
              "text": "Candidate g_009 provides a viable, top-down executable grasp aligned well with the orange cube, with contact center Z at 0.11 m and post-pick lift height at 0.32 m (0.21 m above grasp height), satisfying the transport lift requirement.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Check",
          "title": "Fresh pregrasp check",
          "detail": "A subagent inspects fresh open-hand RGB and the pending pose, then returns continue without applying a nudge.",
          "active_node": "subagent",
          "flow": [
            "backend",
            "subagent"
          ],
          "handoff_label": "Backend \u2192 Subagent \u00b7 fresh pregrasp check",
          "messages": {
            "main": "Grasp the orange cube from above and lift it at least 20 cm for transport.",
            "subagent": "Gripper aligned and open; continue without a nudge.",
            "backend": "Return fresh open-hand RGB and the paused gripper preview."
          },
          "visual": {
            "file": "panda-pick-phase-06.png",
            "original_file": "panda-pick-phase-06.png",
            "label": "Fresh pregrasp check",
            "kind": "Native paused preview",
            "alt": "A subagent inspects fresh open-hand RGB and the pending pose, then returns continue without applying a nudge.",
            "note": "The subagent checks the fresh open-hand scene and returns continue. No nudge is applied."
          },
          "evidence": {
            "event_sequences": [
              41,
              44,
              45,
              47,
              48
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0003.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0003.png",
              "capture_id": "capture_0003",
              "sha256": "fe4c7b235036bb6c79297e8d1e67af511e92c4d43896dde810ad39d42672a1c7"
            }
          ],
          "decision": "continue",
          "nudge_applied": false,
          "candidates": [
            {
              "file": "panda-pick-phase-06.png",
              "label": "Paused mesh",
              "selected": true
            },
            {
              "file": "panda-pick-pregrasp-wrist.png",
              "label": "Fresh wrist RGB",
              "selected": false
            }
          ],
          "start_s": 17.0,
          "end_s": 21.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "pick up the orange cube using a top-down grasp with transport lift height at least 0.20m above grasp",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 47,
              "text": "The open gripper is well-aligned directly over the orange cube with the object centered between the finger pads, appropriate clearance above support, and correct top-down orientation. No nudge required; proceeding to close.",
              "representation": "Recorded original text"
            },
            "runtime_instruction": {
              "sender": "Backend",
              "recipient": "Subagent",
              "event_seq": 44,
              "text": "Inspect the fresh open-hand pregrasp and pending pose. Correct if useful, then explicitly continue to close or abort. All candidate sources allow finite local roll/pitch/yaw without angular magnitude limits; retain translation bounds and validate the changed route. Original delegated goal: pick up the orange cube using a top-down grasp with transport lift height at least 0.20m above grasp",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Lift",
          "title": "Close and lift",
          "detail": "The backend closes and lifts, then returns fresh camera observations and a succeeded command result.",
          "active_node": "backend",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 close and lift",
          "messages": {
            "main": "Grasp the orange cube from above and lift it at least 20 cm for transport.",
            "subagent": "Continue with the inspected grasp.",
            "backend": "Close and lift; return measured robot feedback and fresh images."
          },
          "visual": {
            "file": "panda-pick-phase-07.png",
            "original_file": "panda-pick-phase-07.png",
            "label": "Close and lift",
            "kind": "Recorded RGB",
            "alt": "The backend closes and lifts, then returns fresh camera observations and a succeeded command result.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              49,
              50,
              51,
              52,
              53
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0004.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0004.png",
              "capture_id": "capture_0004",
              "sha256": "358320f773e45706d524860d475478d01cdff2fbaaca4eb5cec0fe6454322378"
            },
            {
              "file": "panda-pick-capture_0005.png",
              "capture_id": "capture_0005",
              "sha256": "db83059a3eaad85c9cc77a3a5592cba4b68728198b42feb3533edc45e0f581fa"
            },
            {
              "file": "panda-pick-capture_0006.png",
              "capture_id": "capture_0006",
              "sha256": "500baa65ec5c97380b9eb49ad1a68972c617d7efdf56ae76c19b985ecf38f05b"
            },
            {
              "file": "panda-pick-capture_0007.png",
              "capture_id": "capture_0007",
              "sha256": "a50401b908fcc3a8eeeef311ce2cbf7dd2a9e946b1bcfc59285bc0c505af1da5"
            }
          ],
          "start_s": 21.0,
          "end_s": 27.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "pick up the orange cube using a top-down grasp with transport lift height at least 0.20m above grasp",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 47,
              "text": "The open gripper is well-aligned directly over the orange cube with the object centered between the finger pads, appropriate clearance above support, and correct top-down orientation. No nudge required; proceeding to close.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Receiver",
          "title": "Select destination",
          "detail": "The main agent asks a subagent to locate the inside bottom of the tray. The backend measures the selected region.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 select destination",
          "messages": {
            "main": "Select the inside bottom of the tray.",
            "subagent": "Tray interior selected and measured.",
            "backend": "Return the selected destination geometry."
          },
          "visual": {
            "file": "panda-pick-destination-focus.jpg",
            "original_file": "panda-pick-phase-08.png",
            "label": "Select destination",
            "kind": "Native point + mask",
            "alt": "The main agent asks a subagent to locate the inside bottom of the tray. The backend measures the selected region.",
            "note": "Original selection cross and segmentation mask, enlarged from the recorded camera view."
          },
          "evidence": {
            "event_sequences": [
              54,
              58,
              59,
              61,
              63,
              64,
              65,
              66
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0007.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0007.png",
              "capture_id": "capture_0007",
              "sha256": "a50401b908fcc3a8eeeef311ce2cbf7dd2a9e946b1bcfc59285bc0c505af1da5"
            }
          ],
          "selection": {
            "normalized_uv": [
              585,
              575
            ],
            "normalized_coordinate_range": [
              0,
              1000
            ],
            "native_marker_pixel_xy": [
              748,
              413
            ],
            "depth_refinement_seed_pixel_xy": [
              744,
              405
            ],
            "observation_id": "obs_0007",
            "point_ref": "pt_014"
          },
          "start_s": 27.0,
          "end_s": 31.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 54,
              "text": "select the inside bottom of the grey/white rectangular tray as the placement destination",
              "representation": "Recorded original text",
              "tool": "delegate_destination"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 63,
              "text": "The segmented region and clicked point accurately identify the inside bottom surface of the grey/white rectangular tray for placement.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Pose",
          "title": "Prepare placement",
          "detail": "The subagent prepares a placement using the tray region: median XY and segment median height plus 0.04 m.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "main"
          ],
          "handoff_label": "Subagent \u2192 Main agent \u00b7 prepare placement",
          "messages": {
            "main": "Place the orange cube in the center bottom of the tray.",
            "subagent": "Placement prepared and validated over the tray center.",
            "backend": "Return the validated native placement preview."
          },
          "visual": {
            "file": "panda-pick-phase-09.png",
            "original_file": "panda-pick-phase-09.png",
            "label": "Prepare placement",
            "kind": "Native placement preview",
            "alt": "The subagent prepares a placement using the tray region: median XY and segment median height plus 0.04 m.",
            "note": "The accepted prepared placement, linked to its original camera projection."
          },
          "evidence": {
            "event_sequences": [
              68,
              73,
              74,
              76,
              77,
              79,
              81
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0007.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0007.png",
              "capture_id": "capture_0007",
              "sha256": "a50401b908fcc3a8eeeef311ce2cbf7dd2a9e946b1bcfc59285bc0c505af1da5"
            }
          ],
          "placement_plan": [
            0.7643049955368042,
            -0.11497705057263374,
            0.13033192321658135
          ],
          "start_s": 31.0,
          "end_s": 35.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 68,
              "text": "place the orange cube into the center bottom of the tray",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 81,
              "text": "Placement pose x1place_025 is prepared and validated directly over the inside center bottom of the tray at transit height 0.315 m down to final height 0.130 m.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Place",
          "title": "Execute placement",
          "detail": "The backend executes the validated placement and returns a fresh observation before release.",
          "active_node": "backend",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 execute placement",
          "messages": {
            "main": "Place the orange cube in the center bottom of the tray.",
            "subagent": "Return the accepted placement for execution.",
            "backend": "Move the held cube to the accepted placement pose."
          },
          "visual": {
            "file": "panda-pick-phase-10.png",
            "original_file": "panda-pick-phase-10.png",
            "label": "Execute placement",
            "kind": "Recorded RGB",
            "alt": "The backend executes the validated placement and returns a fresh observation before release.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              82,
              83,
              84,
              85
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0008.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0008.png",
              "capture_id": "capture_0008",
              "sha256": "b7a4c7898e7970bbb1108d25eae40ffb9d21078a05ae21ee47b06e3a64aaa49d"
            }
          ],
          "start_s": 35.0,
          "end_s": 38.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 68,
              "text": "place the orange cube into the center bottom of the tray",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 81,
              "text": "Placement pose x1place_025 is prepared and validated directly over the inside center bottom of the tray at transit height 0.315 m down to final height 0.130 m.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Release",
          "title": "Release",
          "detail": "The main agent opens the gripper and commands a 0.15 m upward retreat; fresh observations follow.",
          "active_node": "main",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 release",
          "messages": {
            "main": "Place the orange cube in the center bottom of the tray.",
            "subagent": "The accepted placement is ready for release.",
            "backend": "Open the gripper and return fresh camera observations."
          },
          "visual": {
            "file": "panda-pick-phase-11.png",
            "original_file": "panda-pick-phase-11.png",
            "label": "Release",
            "kind": "Recorded RGB",
            "alt": "The main agent opens the gripper and commands a 0.15 m upward retreat; fresh observations follow.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              86,
              87,
              88,
              89,
              91,
              92
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0009.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0009.png",
              "capture_id": "capture_0009",
              "sha256": "6c7f01ca951fea2acdb79c251243effc442970bbd5327ca433c13d86a47607be"
            },
            {
              "file": "panda-pick-capture_0010.png",
              "capture_id": "capture_0010",
              "sha256": "293343b51d0ffc331939d44ffdbcb740c1c065a122cdc7e832af386c9aeba82a"
            },
            {
              "file": "panda-pick-capture_0011.png",
              "capture_id": "capture_0011",
              "sha256": "8bbad2754659f90265425e79c5fc3405f3b6984215641acb9c907d5f8c28268b"
            }
          ],
          "start_s": 38.0,
          "end_s": 42.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 68,
              "text": "place the orange cube into the center bottom of the tray",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 81,
              "text": "Placement pose x1place_025 is prepared and validated directly over the inside center bottom of the tray at transit height 0.315 m down to final height 0.130 m.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Outcome",
          "title": "Outcome observation",
          "detail": "Return the arm home and report completion from fresh camera images.",
          "active_node": "main",
          "flow": [
            "backend",
            "main"
          ],
          "handoff_label": "Backend \u2192 Main agent \u00b7 outcome observation",
          "messages": {
            "main": "Place the orange cube in the center bottom of the tray.",
            "subagent": "Placement pose prepared and validated for execution.",
            "backend": "Return the final camera observation after retreat."
          },
          "visual": {
            "file": "panda-pick-phase-12.png",
            "original_file": "panda-pick-phase-12.png",
            "label": "Outcome observation",
            "kind": "Recorded RGB",
            "alt": "The main agent returns the arm home and reports completion from fresh camera images. The orange cube is visible inside the tray.",
            "note": "The final camera image shows the requested object relation. Completion is agent-reported; no independent physical task verifier is available."
          },
          "evidence": {
            "event_sequences": [
              94,
              95,
              96,
              97
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-pick-capture_0012.png",
          "camera_sequence": [
            {
              "file": "panda-pick-capture_0012.png",
              "capture_id": "capture_0012",
              "sha256": "c948a11be7daff7670db518fb722c0d9fd3c8f4e1d522e3498395273daed0420"
            }
          ],
          "start_s": 42.0,
          "end_s": 45.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 68,
              "text": "place the orange cube into the center bottom of the tray",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 81,
              "text": "Placement pose x1place_025 is prepared and validated directly over the inside center bottom of the tray at transit height 0.315 m down to final height 0.130 m.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Last instruction \u2192 Subagent",
            "subagent": "Latest report \u2192 Main agent",
            "backend": "Final camera observation"
          }
        }
      ],
      "presentation_duration_s": 45.0,
      "language_source_sha256": "0cad67327717163a4e381caa6f5eca822b15dcfaaf817161cfb6757ac749203b"
    },
    {
      "id": "panda-stack-cubes",
      "label": "Panda \u00b7 Stack cubes",
      "task": "Stack the orange cube on top of the green cube.",
      "episode_id": "gui_20260926_121447_90cfdaad",
      "video_file": "panda-stack-cubes-motion.mp4",
      "demo_file": "panda-stack-cubes-demo.mp4",
      "poster_file": "panda-stack-capture_0001.png",
      "motion_label": "Recorded camera observations",
      "timing_note": "Camera observations paired with native tool images. Recorded points and gripper poses are shown alongside the physical scene.",
      "camera_frames": [
        {
          "file": "panda-stack-capture_0001.png",
          "capture_id": "capture_0001",
          "sha256": "f69fbaa02c8c6626a0a751ac1bf7909e7e37117b9d61bd98307ec616137e74dc"
        },
        {
          "file": "panda-stack-capture_0002.png",
          "capture_id": "capture_0002",
          "sha256": "2e11cfbdf49d50ab1ab80bffc16d5158d7244451a617a6c1b5d8f4ef54fcfbef"
        },
        {
          "file": "panda-stack-capture_0003.png",
          "capture_id": "capture_0003",
          "sha256": "fc5d3651eced39a4684a9bb9895cb8202d07995d2890822836092804407dc1b9"
        },
        {
          "file": "panda-stack-capture_0004.png",
          "capture_id": "capture_0004",
          "sha256": "39b16c67acd7ec6fa36c37f7f970aad0e441648e0e1f2d81cdfa5144113b1198"
        },
        {
          "file": "panda-stack-capture_0005.png",
          "capture_id": "capture_0005",
          "sha256": "17735454f57e7d022020f67ffda3acb24dca82f736726d6134dad757100beef8"
        },
        {
          "file": "panda-stack-capture_0006.png",
          "capture_id": "capture_0006",
          "sha256": "219d5731a7151f46b50c02a0d4717a56b003c3be1bf99028a09a14a907f67a08"
        },
        {
          "file": "panda-stack-capture_0007.png",
          "capture_id": "capture_0007",
          "sha256": "d1dc885da448439c755d66e01ed5c4c882c9c2020d55ce1641e4131b47f4b753"
        },
        {
          "file": "panda-stack-capture_0008.png",
          "capture_id": "capture_0008",
          "sha256": "7a78f730b0f7921bf49df2aed80ba78621874ffa836027b1a38029eac9877f5f"
        },
        {
          "file": "panda-stack-capture_0009.png",
          "capture_id": "capture_0009",
          "sha256": "ec515503ec45dca99195bc2c96f55c4b9ea754db58889cbb8a88f8c00b4f7a6c"
        },
        {
          "file": "panda-stack-capture_0010.png",
          "capture_id": "capture_0010",
          "sha256": "be4b26690ee0c1af4689967679a89bdd963485a58a18e211d51dfb4a6a4e719c"
        },
        {
          "file": "panda-stack-capture_0011.png",
          "capture_id": "capture_0011",
          "sha256": "6a4363cdea582dab28680a8aba89da4ba2372ea1dcdc6cde54a17e77c2c9b66b"
        }
      ],
      "outcome": {
        "agent_status": "completed",
        "agent_verified": false,
        "physical_task_success_flag": false,
        "display_label": "Agent-reported completion; final camera observation shown",
        "independent_physical_success_verifier": false
      },
      "selection_note": "Compact accepted sequence; earlier proposal attempts and tool-contract retries omitted. All displayed tool images and camera samples come from the same execution.",
      "grasp_source": "median",
      "selected_candidate": "g_005",
      "phases": [
        {
          "short": "Request",
          "title": "Main issues a subgoal",
          "detail": "Select the orange cube in the current image.",
          "active_node": "main",
          "flow": [
            "main",
            "subagent"
          ],
          "handoff_label": "Main agent \u2192 Subagent \u00b7 request the visual target",
          "messages": {
            "main": "Select the orange cube in the current image.",
            "subagent": "Point to the cube and return the measured region.",
            "backend": "Return the current robot camera views."
          },
          "visual": {
            "file": "panda-stack-phase-01.png",
            "original_file": "panda-stack-phase-01.png",
            "label": "Language goal",
            "kind": "Recorded RGB",
            "alt": "The main agent asks a subagent to locate the orange cube.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              3,
              4
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "f69fbaa02c8c6626a0a751ac1bf7909e7e37117b9d61bd98307ec616137e74dc"
            }
          ],
          "start_s": 0,
          "end_s": 2.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 4,
              "text": "Click on the orange cube to grasp it.",
              "representation": "Recorded original text",
              "tool": "delegate_point"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Point",
          "title": "Point to measured mask",
          "detail": "The subagent chooses a point on the orange cube. The backend returns the measured SAM region.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 point to measured mask",
          "messages": {
            "main": "Select the orange cube in the current image.",
            "subagent": "Orange cube selected; mask and measured geometry returned.",
            "backend": "Return the selected point, segmentation mask and 3D geometry."
          },
          "visual": {
            "file": "panda-stack-target-focus.jpg",
            "original_file": "panda-stack-phase-02.png",
            "label": "Point to measured mask",
            "kind": "Native point + mask",
            "alt": "The subagent chooses a point on the orange cube. The backend returns the measured SAM region.",
            "note": "Original selection cross and segmentation mask, enlarged from the recorded camera view."
          },
          "evidence": {
            "event_sequences": [
              8,
              9,
              11,
              13
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "f69fbaa02c8c6626a0a751ac1bf7909e7e37117b9d61bd98307ec616137e74dc"
            }
          ],
          "selection": {
            "normalized_uv": [
              468,
              635
            ],
            "normalized_coordinate_range": [
              0,
              1000
            ],
            "native_marker_pixel_xy": [
              599,
              457
            ],
            "depth_refinement_seed_pixel_xy": [
              599,
              459
            ],
            "observation_id": "obs_0001",
            "point_ref": "pt_002"
          },
          "start_s": 2.0,
          "end_s": 6.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 4,
              "text": "Click on the orange cube to grasp it.",
              "representation": "Recorded original text",
              "tool": "delegate_point"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 13,
              "text": "Successfully selected the orange cube with clean segmentation and calibrated geometry.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Grasps",
          "title": "Grasp candidates",
          "detail": "The backend returns three feasible geometric median grasp candidates after the vertical proposal is rejected.",
          "active_node": "backend",
          "flow": [
            "backend",
            "subagent"
          ],
          "handoff_label": "Backend \u2192 Subagent \u00b7 grasp candidates",
          "messages": {
            "main": "Grasp the orange cube and lift it at least 20 cm for transport.",
            "subagent": "Compare 3 feasible geometric gripper poses.",
            "backend": "Return 3 geometric median grasp proposals."
          },
          "visual": {
            "file": "panda-stack-g_005.png",
            "original_file": "panda-stack-g_005.png",
            "label": "Grasp candidates",
            "kind": "Native mesh projections",
            "alt": "The backend returns three feasible geometric median grasp candidates after the vertical proposal is rejected.",
            "note": "3 feasible geometric median proposals. g_005 is the recorded selection. Preview occlusion is not tested. Use the thumbnails below to inspect each original proposal."
          },
          "evidence": {
            "event_sequences": [
              16,
              20,
              21,
              23,
              24
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "f69fbaa02c8c6626a0a751ac1bf7909e7e37117b9d61bd98307ec616137e74dc"
            }
          ],
          "candidates": [
            {
              "file": "panda-stack-g_004.png",
              "label": "g_004",
              "selected": false,
              "yaw_deg": -45.0,
              "source": "median",
              "contact_center_xyz_m": [
                0.6129314219951629,
                0.0173236451111734,
                0.10416273772716521
              ],
              "transit": {
                "pre_pick_z_m": 0.3541627377271652,
                "post_pick_z_m": 0.3541627377271652
              }
            },
            {
              "file": "panda-stack-g_005.png",
              "label": "g_005 \u00b7 selected",
              "selected": true,
              "yaw_deg": 0.0,
              "source": "median",
              "contact_center_xyz_m": [
                0.6129314219951629,
                0.0173236451111734,
                0.10416273772716521
              ],
              "transit": {
                "pre_pick_z_m": 0.3541627377271652,
                "post_pick_z_m": 0.3541627377271652
              }
            },
            {
              "file": "panda-stack-g_006.png",
              "label": "g_006",
              "selected": false,
              "yaw_deg": 45.0,
              "source": "median",
              "contact_center_xyz_m": [
                0.6129314219951629,
                0.0173236451111734,
                0.10416273772716521
              ],
              "transit": {
                "pre_pick_z_m": 0.3541627377271652,
                "post_pick_z_m": 0.3541627377271652
              }
            }
          ],
          "start_s": 6.0,
          "end_s": 10.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "Grasp the orange cube for transport. Ensure outgoing height is at least 0.20 m above grasp height.",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Visual choice in progress",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Inspect",
          "title": "Inspect and choose",
          "detail": "The subagent inspects g_005 and returns this top-down grasp, with incoming and outgoing heights 0.25 m above contact.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "main"
          ],
          "handoff_label": "Subagent \u2192 Main agent \u00b7 inspect and choose",
          "messages": {
            "main": "Grasp the orange cube and lift it at least 20 cm for transport.",
            "subagent": "Choose g_005: top-down grasp with sufficient transport lift.",
            "backend": "Return native gripper projections and point-cloud inspection views."
          },
          "visual": {
            "file": "panda-stack-grasp-inspection.jpg",
            "original_file": "panda-stack-phase-04.png",
            "label": "Inspect and choose",
            "kind": "Native point-cloud view",
            "alt": "The subagent inspects g_005 and returns this top-down grasp, with incoming and outgoing heights 0.25 m above contact.",
            "note": "SIDE / TOP / CLOSING PLANE panels of g_005. Opening: 80 mm. Lower rotation examples are explanatory; no pose edit was applied."
          },
          "evidence": {
            "event_sequences": [
              26,
              27,
              29,
              31,
              33
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0001.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0001.png",
              "capture_id": "capture_0001",
              "sha256": "f69fbaa02c8c6626a0a751ac1bf7909e7e37117b9d61bd98307ec616137e74dc"
            }
          ],
          "selected_candidate": "g_005",
          "nominal_opening_mm": 80,
          "start_s": 10.0,
          "end_s": 14.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "Grasp the orange cube for transport. Ensure outgoing height is at least 0.20 m above grasp height.",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 33,
              "text": "Candidate g_005 is a top-down grasp centered on the orange cube with transit pre and post heights configured to 0.25 m above grasp, fulfilling the transport requirement.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Approach",
          "title": "Approach",
          "detail": "The backend validates g_005 and starts the physical grasp primitive.",
          "active_node": "backend",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 approach",
          "messages": {
            "main": "Grasp the orange cube and lift it at least 20 cm for transport.",
            "subagent": "Return g_005 for validated physical execution.",
            "backend": "Validate g_005 and start the grasp primitive."
          },
          "visual": {
            "file": "panda-stack-g_005.png",
            "original_file": "panda-stack-g_005.png",
            "label": "Approach",
            "kind": "Native selected pose",
            "alt": "The backend validates g_005 and starts the physical grasp primitive.",
            "note": "The backend validates g_005 and starts the physical grasp primitive."
          },
          "evidence": {
            "event_sequences": [
              34,
              35,
              36
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0002.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0002.png",
              "capture_id": "capture_0002",
              "sha256": "2e11cfbdf49d50ab1ab80bffc16d5158d7244451a617a6c1b5d8f4ef54fcfbef"
            }
          ],
          "start_s": 14.0,
          "end_s": 17.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "Grasp the orange cube for transport. Ensure outgoing height is at least 0.20 m above grasp height.",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 33,
              "text": "Candidate g_005 is a top-down grasp centered on the orange cube with transit pre and post heights configured to 0.25 m above grasp, fulfilling the transport requirement.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Check",
          "title": "Fresh pregrasp check",
          "detail": "A subagent inspects fresh open-hand RGB and the pending pose, then returns continue without applying a nudge.",
          "active_node": "subagent",
          "flow": [
            "backend",
            "subagent"
          ],
          "handoff_label": "Backend \u2192 Subagent \u00b7 fresh pregrasp check",
          "messages": {
            "main": "Grasp the orange cube and lift it at least 20 cm for transport.",
            "subagent": "Gripper aligned and open; continue without a nudge.",
            "backend": "Return fresh open-hand RGB and the paused gripper preview."
          },
          "visual": {
            "file": "panda-stack-phase-06.png",
            "original_file": "panda-stack-phase-06.png",
            "label": "Fresh pregrasp check",
            "kind": "Native paused preview",
            "alt": "A subagent inspects fresh open-hand RGB and the pending pose, then returns continue without applying a nudge.",
            "note": "The subagent checks the fresh open-hand scene and returns continue. No nudge is applied."
          },
          "evidence": {
            "event_sequences": [
              37,
              40,
              41,
              43,
              44
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0003.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0003.png",
              "capture_id": "capture_0003",
              "sha256": "fc5d3651eced39a4684a9bb9895cb8202d07995d2890822836092804407dc1b9"
            }
          ],
          "decision": "continue",
          "nudge_applied": false,
          "candidates": [
            {
              "file": "panda-stack-phase-06.png",
              "label": "Paused mesh",
              "selected": true
            },
            {
              "file": "panda-stack-pregrasp-wrist.png",
              "label": "Fresh wrist RGB",
              "selected": false
            }
          ],
          "start_s": 17.0,
          "end_s": 21.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "Grasp the orange cube for transport. Ensure outgoing height is at least 0.20 m above grasp height.",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 43,
              "text": "Gripper is centered above the orange cube with jaws aligned and open sufficiently; ready to close.",
              "representation": "Recorded original text"
            },
            "runtime_instruction": {
              "sender": "Backend",
              "recipient": "Subagent",
              "event_seq": 40,
              "text": "Inspect the fresh open-hand pregrasp and pending pose. Correct if useful, then explicitly continue to close or abort. All candidate sources allow finite local roll/pitch/yaw without angular magnitude limits; retain translation bounds and validate the changed route. Original delegated goal: Grasp the orange cube for transport. Ensure outgoing height is at least 0.20 m above grasp height.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Lift",
          "title": "Close and lift",
          "detail": "The backend closes and lifts, then returns fresh camera observations and a succeeded command result.",
          "active_node": "backend",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 close and lift",
          "messages": {
            "main": "Grasp the orange cube and lift it at least 20 cm for transport.",
            "subagent": "Continue with the inspected grasp.",
            "backend": "Close and lift; return measured robot feedback and fresh images."
          },
          "visual": {
            "file": "panda-stack-phase-07.png",
            "original_file": "panda-stack-phase-07.png",
            "label": "Close and lift",
            "kind": "Recorded RGB",
            "alt": "The backend closes and lifts, then returns fresh camera observations and a succeeded command result.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              45,
              46,
              47,
              48,
              49
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0004.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0004.png",
              "capture_id": "capture_0004",
              "sha256": "39b16c67acd7ec6fa36c37f7f970aad0e441648e0e1f2d81cdfa5144113b1198"
            },
            {
              "file": "panda-stack-capture_0005.png",
              "capture_id": "capture_0005",
              "sha256": "17735454f57e7d022020f67ffda3acb24dca82f736726d6134dad757100beef8"
            },
            {
              "file": "panda-stack-capture_0006.png",
              "capture_id": "capture_0006",
              "sha256": "219d5731a7151f46b50c02a0d4717a56b003c3be1bf99028a09a14a907f67a08"
            },
            {
              "file": "panda-stack-capture_0007.png",
              "capture_id": "capture_0007",
              "sha256": "d1dc885da448439c755d66e01ed5c4c882c9c2020d55ce1641e4131b47f4b753"
            }
          ],
          "start_s": 21.0,
          "end_s": 27.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 16,
              "text": "Grasp the orange cube for transport. Ensure outgoing height is at least 0.20 m above grasp height.",
              "representation": "Recorded original text",
              "tool": "delegate_grasp"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 43,
              "text": "Gripper is centered above the orange cube with jaws aligned and open sufficiently; ready to close.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Receiver",
          "title": "Select destination",
          "detail": "The main agent asks a subagent to locate the top of the green cube. The backend measures the selected region.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "backend"
          ],
          "handoff_label": "Subagent \u2192 Backend \u00b7 select destination",
          "messages": {
            "main": "Select the top surface of the green cube.",
            "subagent": "Green cube top selected as the destination.",
            "backend": "Return the selected destination geometry."
          },
          "visual": {
            "file": "panda-stack-destination-focus.jpg",
            "original_file": "panda-stack-phase-08.png",
            "label": "Select destination",
            "kind": "Native point + mask",
            "alt": "The main agent asks a subagent to locate the top of the green cube. The backend measures the selected region.",
            "note": "Original selection cross and segmentation mask, enlarged from the recorded camera view."
          },
          "evidence": {
            "event_sequences": [
              50,
              54,
              55,
              57,
              59,
              60,
              61,
              62
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0007.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0007.png",
              "capture_id": "capture_0007",
              "sha256": "d1dc885da448439c755d66e01ed5c4c882c9c2020d55ce1641e4131b47f4b753"
            }
          ],
          "selection": {
            "normalized_uv": [
              654,
              717
            ],
            "normalized_coordinate_range": [
              0,
              1000
            ],
            "native_marker_pixel_xy": [
              836,
              516
            ],
            "depth_refinement_seed_pixel_xy": [
              836,
              516
            ],
            "observation_id": "obs_0007",
            "point_ref": "pt_010"
          },
          "start_s": 27.0,
          "end_s": 31.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 50,
              "text": "Click on the top surface of the green cube to place the orange cube on it.",
              "representation": "Recorded original text",
              "tool": "delegate_destination"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 59,
              "text": "Successfully selected the top surface of the green cube as the placement destination.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Pose",
          "title": "Prepare placement",
          "detail": "The subagent prepares a placement using the green cube region: median XY and observed maximum height plus 0.04 m.",
          "active_node": "subagent",
          "flow": [
            "subagent",
            "main"
          ],
          "handoff_label": "Subagent \u2192 Main agent \u00b7 prepare placement",
          "messages": {
            "main": "Place the held orange cube on top of the green cube.",
            "subagent": "Placement prepared and validated on the green cube.",
            "backend": "Return the validated native placement preview."
          },
          "visual": {
            "file": "panda-stack-phase-09.png",
            "original_file": "panda-stack-phase-09.png",
            "label": "Prepare placement",
            "kind": "Native placement preview",
            "alt": "The subagent prepares a placement using the green cube region: median XY and observed maximum height plus 0.04 m.",
            "note": "The accepted prepared placement, linked to its original camera projection."
          },
          "evidence": {
            "event_sequences": [
              64,
              69,
              70,
              72,
              73,
              75,
              77
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0007.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0007.png",
              "capture_id": "capture_0007",
              "sha256": "d1dc885da448439c755d66e01ed5c4c882c9c2020d55ce1641e4131b47f4b753"
            }
          ],
          "placement_plan": [
            0.49527250245213505,
            -0.25663151405751705,
            0.16449765205383302
          ],
          "start_s": 31.0,
          "end_s": 35.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 64,
              "text": "Place the held orange cube on top of the green cube.",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 77,
              "text": "Placement pose planned and validated directly on top of the green cube at target contact center height Z=0.164m with clear transit path.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Place",
          "title": "Execute placement",
          "detail": "The backend executes the validated placement and returns a fresh observation before release.",
          "active_node": "backend",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 execute placement",
          "messages": {
            "main": "Place the held orange cube on top of the green cube.",
            "subagent": "Return the accepted placement for execution.",
            "backend": "Move the held cube to the accepted placement pose."
          },
          "visual": {
            "file": "panda-stack-phase-10.png",
            "original_file": "panda-stack-phase-10.png",
            "label": "Execute placement",
            "kind": "Recorded RGB",
            "alt": "The backend executes the validated placement and returns a fresh observation before release.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              78,
              79,
              80,
              81
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0008.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0008.png",
              "capture_id": "capture_0008",
              "sha256": "7a78f730b0f7921bf49df2aed80ba78621874ffa836027b1a38029eac9877f5f"
            }
          ],
          "start_s": 35.0,
          "end_s": 38.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 64,
              "text": "Place the held orange cube on top of the green cube.",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 77,
              "text": "Placement pose planned and validated directly on top of the green cube at target contact center height Z=0.164m with clear transit path.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Release",
          "title": "Release",
          "detail": "The main agent opens the gripper; the backend returns fresh observations after release.",
          "active_node": "main",
          "flow": [
            "main",
            "backend"
          ],
          "handoff_label": "Main agent \u2192 Backend \u00b7 release",
          "messages": {
            "main": "Place the held orange cube on top of the green cube.",
            "subagent": "The accepted placement is ready for release.",
            "backend": "Open the gripper and return fresh camera observations."
          },
          "visual": {
            "file": "panda-stack-phase-11.png",
            "original_file": "panda-stack-phase-11.png",
            "label": "Release",
            "kind": "Recorded RGB",
            "alt": "The main agent opens the gripper; the backend returns fresh observations after release.",
            "note": "Original camera observation from this physical execution."
          },
          "evidence": {
            "event_sequences": [
              82,
              83,
              84,
              85
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0009.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0009.png",
              "capture_id": "capture_0009",
              "sha256": "ec515503ec45dca99195bc2c96f55c4b9ea754db58889cbb8a88f8c00b4f7a6c"
            },
            {
              "file": "panda-stack-capture_0010.png",
              "capture_id": "capture_0010",
              "sha256": "be4b26690ee0c1af4689967679a89bdd963485a58a18e211d51dfb4a6a4e719c"
            }
          ],
          "start_s": 38.0,
          "end_s": 42.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 64,
              "text": "Place the held orange cube on top of the green cube.",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 77,
              "text": "Placement pose planned and validated directly on top of the green cube at target contact center height Z=0.164m with clear transit path.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Sent instruction \u2192 Subagent",
            "subagent": "Report \u2192 Main agent",
            "backend": "Visual feedback & execution"
          }
        },
        {
          "short": "Outcome",
          "title": "Outcome observation",
          "detail": "Retreat upward by 0.15 m and report completion from fresh camera images.",
          "active_node": "main",
          "flow": [
            "backend",
            "main"
          ],
          "handoff_label": "Backend \u2192 Main agent \u00b7 outcome observation",
          "messages": {
            "main": "Place the held orange cube on top of the green cube.",
            "subagent": "Placement pose prepared and validated for execution.",
            "backend": "Return the final camera observation after retreat."
          },
          "visual": {
            "file": "panda-stack-phase-12.png",
            "original_file": "panda-stack-phase-12.png",
            "label": "Outcome observation",
            "kind": "Recorded RGB",
            "alt": "The main agent commands a 0.15 m upward retreat and reports completion from fresh camera images. The orange cube is visible on the green cube.",
            "note": "The final camera image shows the requested object relation. Completion is agent-reported; no independent physical task verifier is available."
          },
          "evidence": {
            "event_sequences": [
              87,
              88,
              89,
              90
            ]
          },
          "source_start_s": 0,
          "source_end_s": 0,
          "camera_file": "panda-stack-capture_0011.png",
          "camera_sequence": [
            {
              "file": "panda-stack-capture_0011.png",
              "capture_id": "capture_0011",
              "sha256": "6a4363cdea582dab28680a8aba89da4ba2372ea1dcdc6cde54a17e77c2c9b66b"
            }
          ],
          "start_s": 42.0,
          "end_s": 45.0,
          "language_handoff": {
            "request": {
              "sender": "Main agent",
              "recipient": "Subagent",
              "event_seq": 64,
              "text": "Place the held orange cube on top of the green cube.",
              "representation": "Recorded original text",
              "tool": "delegate_place"
            },
            "report": {
              "sender": "Subagent",
              "recipient": "Main agent",
              "event_seq": 77,
              "text": "Placement pose planned and validated directly on top of the green cube at target contact center height Z=0.164m with clear transit path.",
              "representation": "Recorded original text"
            }
          },
          "node_labels": {
            "main": "Last instruction \u2192 Subagent",
            "subagent": "Latest report \u2192 Main agent",
            "backend": "Final camera observation"
          }
        }
      ],
      "presentation_duration_s": 45.0,
      "language_source_sha256": "6cdce061c7c058155b8ec919b3e9569685a85fce1f10354ff3dad7175765408c"
    }
  ],
  "language_presentation_note": "The Main instruction remains visible while the Subagent chooses, inspects and returns visual evidence. Recorded request/return messages are grouped within the selected reading intervals; presentation times are not event wall-clock times."
}