EM3D-GEO-1 re-export: 32 episodes @15fps , 512x384 portrait obs, native depth on every episode, metric world state; supersedes the July 24-episode drop

README.md

@@ -32,7 +32,7 @@ Recorded on iPhone Pro (ARKit LiDAR) as part of Everyday Manipulation 1 (EM1),
 annotated by CaryX AI, packaged as a single multi-episode
 [LeRobot v3.0](https://github.com/huggingface/lerobot) dataset.
 
-**23 episodes · 4393 frames @ 10 fps · tasks: fold_laundry, jar_open_close, scoop_grain**
+**32 episodes · 6865 frames @ 15 fps · tasks: fold_laundry, jar_open_close, scoop_grain, sweep**
 
 ## Quick start
 
@@ -49,24 +49,34 @@ frame = ds[0]  # RGB, metric depth, MANO hands, contact, camera pose in one dict
 **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC-BY-NC-SA-4.0)**, research use only.
 https://creativecommons.org/licenses/by-nc-sa/4.0/. Full terms in `LICENSE.txt`.
 
+## Episode orientation
+
+Most episodes were recorded portrait; a few were recorded landscape and are
+shipped ROTATED 90° CCW onto the same portrait canvas so the dataset merges
+under one schema. Those episodes are flagged in
+`info.foundry.episodes.<id>.raster_rotation = "ccw90"`: rotate their image
+channels 90° CW to display upright. All spatial channels (RGB, depth, masks,
+camera pose, MANO orientation) are expressed consistently in the shipped,
+rotated frame, so training and 3D geometry need no special-casing.
+
 ## Features
 
 Shapes are numpy shapes, so `(7,)` is a flat 7-element vector.
 
 | key | shape | notes |
 |---|---|---|
-| `observation.images.ego` | (256, 256, 3) | RGB ego frame |
-| `observation.images.depth` | (192-h, w, 1) video | **metric metres**, 12-bit log-HEVC (quantizer [0.1, 5.0] m in `meta/info.json`). Depth has its OWN intrinsics, listed in `episodes/<id>/depth_geometry.json`. Never assume `depth[y,x] == ego[y,x]`. |
-| `observation.state` | (7,) | wrist pose `[x, y, z, rx, ry, rz]` plus grasp. Frame is per-episode `action_frame` in `info.foundry.episodes`. |
+| `observation.images.ego` | (512, 384, 3) | RGB ego frame, portrait, ONE uniform downscale of the native 1440×1920 (no aspect distortion). Exactly 2× the depth grid: `ego[y, x]` and `depth[y//2, x//2]` are the same ray. |
+| `observation.images.depth` | (256, 192, 1) video | **metric metres**, native ARKit sceneDepth grid (never resampled), 12-bit log-HEVC (quantizer [0.1, 5.0] m in `meta/info.json`). Intrinsics for both grids: `episodes/<id>/depth_geometry.json`. |
+| `observation.state` | (7,) | wrist pose `[x, y, z, rx, ry, rz]` plus grasp, metric WORLD frame. The wrist source is named per episode in `info.foundry.episodes.state_source` (LiDAR-solved joints, or the RTMPose wrist depth-lifted through the LiDAR grid). |
 | `action` | (7,) | wrist pose delta `[dx, dy, dz, drx, dry, drz]` plus grasp. |
-| `observation.camera_pose` | (4, 4) | ARKit camera to world (metric) |
+| `observation.camera_pose` | (4, 4) | camera to world (metric), expressed in the SHIPPED raster frame (x = raster right, y = raster down, z = forward) and gravity-verified per episode at export. Composes directly with the intrinsics in `depth_geometry.json`. |
 | `observation.contact` | (2, 2) | The shipped contact channel. Rows are the left and right hand, column k is object slot k. Values are contact strength in [0, 1]. Contact against any object is the max over slots. |
 | `observation.contact_valid` | (2, 2) | 1.0 where that hand-to-object distance was actually measured on that frame, 0.0 where it could not be (hand out of frame, no depth surface, or no object in that slot). Single-object episodes ship slot 1 all-zero rather than a fabricated "measured apart". |
 | `observation.mano_joints` | (2, 21, 3) | **ROOT-RELATIVE** MANO/HaMeR joints (wrist at the origin). Compose with `observation.state` + `observation.camera_pose` for world placement. The primary channel is tip-smoothed (occlusion-gated fingertip de-jitter, named per episode in `mano_joints_channel`); the raw fit is recoverable from `observation.mano_params`. |
 | `observation.mano_params` | (2, 58) | full MANO parameterization: global_orient(3) ⊕ pose(45) ⊕ betas(10) |
 | `observation.mano_valid` | (2,) | 1.0 where the MANO fit is valid for that hand/frame |
-| `observation.images.object_mask` | (256, 256, 3) video | R = object slot 0, B = slot 1 (two-object episodes), G unused. Threshold > 127 (video codecs are lossy). Exact full-res masks: `episodes/<id>/object_mask.npz`. |
-| `observation.images.hand_mask` | (256, 256, 3) video | R = LEFT hand, B = RIGHT hand. Same thresholding. Reviewer-corrected (shipped) channel. |
+| `observation.images.object_mask` | (512, 384, 3) video | R = object slot 0, B = slot 1 (two-object episodes), G unused. Threshold > 127 (video codecs are lossy). Exact full-res masks: `episodes/<id>/object_mask.npz`. |
+| `observation.images.hand_mask` | (512, 384, 3) video | R = LEFT hand, B = RIGHT hand. Same thresholding. Reviewer-corrected (shipped) channel. |
 
 ## Per-episode sidecars
 

data/chunk-000/file-000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d5cd0639218e9b695f26b45373fcff2a43d68446d343248065813454931cc3e2
-size 94132875
+oid sha256:80d4ed4fa9bb18297ed75e1582e660d43a28ff916d79e1d6a7c999817ff3183e
+size 61109349

data/chunk-000/file-001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:178d6559059af4c06b62a8d8ecc543f14644bd100eaad7b7b4d51ec412f23a68
-size 102366693
+oid sha256:fefce740534ac67ed7568c4829382654cc16c141b0516cc91dd99890685eac58
+size 61526406

data/chunk-000/file-002.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:28580caee7ea6568ee198c1cd2a257865281ef92b21d080885dfd4693f337bc2
-size 103206450
+oid sha256:4fdb5915fa71da066f1125fbf4a0cf4ad1aa2ffaa7eec53c5511dc5c621f9038
+size 58618948

data/chunk-000/file-003.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:88301ca026f275b9a06ebde61bbcd3cc3159ce51ec11d5e932ca9d1c5ce3daa3
-size 83992360
+oid sha256:6446096156b19610cf871f998346a75f11cb67629d8e5656896b0671dc404838
+size 87774756

data/chunk-000/file-004.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f1a5c6636a6fd1c9c50e95787a3459d06b533c160319b4e408ecd41f69bc9125
-size 44725729
+oid sha256:9598cffce5b2a6294ad043c6b488e69eeb18ec9cf863fa227f2386cd64171709
+size 57112144

data/chunk-000/file-005.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a1fcfd52b9ce8493fd014347380a19c908179dbc844ce7fa7fb10e69a3805b6
+size 80945220

data/chunk-000/file-006.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35b150e44219333d35753a688003301c3fad840d8fba08b2faa132c7a9cf3c13
+size 83043577

data/chunk-000/file-007.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fe94432c433c3c2e343b6afc3e70c36cb742912530dfa1efcd23248668896a6
+size 74293222

data/chunk-000/file-008.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da24e93427ca585529ffdf9b6efdb7108ead99148a7cc6c53f3132a4a1661792
+size 67108426

data/chunk-000/file-009.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:470663947d9d00f83def30a99fc48f0728453c8932c5776ac55c447f0770f13d
+size 81826773

data/chunk-000/file-010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac019cf4aaf5b02d6c39844523239a97905b483040d9ade91f7d94db3ef78b23
+size 80064331

data/chunk-000/file-011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2c939eb7293fdbafe737e03a3133a4c6e35b631191f7d2a8e0de93f46ec82d3
+size 104299672

data/chunk-000/file-012.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b7fab3586be748ae138476d889b7f58a75ea437b189ef4cb14b83deb7559ceb
+size 69607465

data/chunk-000/file-013.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24a5b8d53a99ac536e0d48cfbf4cb781c8c85fc0030fd8c04f91cde4dc2f7370
+size 117634762

data/chunk-000/file-014.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b20196296977dafa62748ef2316c72c18acc78c974105c537b115b578492849
+size 70280940

data/chunk-000/file-015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7dcac3d086c8f1bcd7171e3f7f0b8a5260514d5c320c291ecf335a9961497545
+size 65395115

data/chunk-000/file-016.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1646c8dc2ff4c0e414f6e432bb6882a0a2268008f9d85641f0ae473ff35af9ed
+size 101807056

data/chunk-000/file-017.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67307bf311dcd21b3c76aa21938db1c2ad3a9bab61cb4492142e11a63858614b
+size 81084115

data/chunk-000/file-018.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4fe6538db945c75d1e42245bd767724da7f32bccf342bdfc772c7666a55f02d
+size 80221418

data/chunk-000/file-019.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:675c793b2ef2408261dabad54b93044b37227700fc7b0bcccfb86a9ce81e8fdb
+size 91390242

data/chunk-000/file-020.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4e2e92f06cce0cb4b38bd35603d494a243e3e223e02d95d7571812ac35d08bb
+size 91679790

data/chunk-000/file-021.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adf827b5d8e29cfab05ebc2096c4741bc51281385409ee6d4237bd626f6337e6
+size 95456727

episodes/019fb6df-0a15-7630-b4e1-b5e552791fcc/depth_geometry.json

@@ -1,14 +1,18 @@
 {
   "sensor": "iphone_lidar_arkit",
-  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth, its OWN intrinsics \u2014 see depth_geometry.json); observation.images.ego is 256x256 (resized). DIFFERENT pixel grids \u2014 register via the depth intrinsics + observation.camera_pose, never depth[y,x]==ego[y,x]. The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
   "depth_resolution_hw": [
     256,
     192
   ],
   "ego_obs_resolution_hw": [
-    256,
-    256
+    512,
+    384
   ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": null,
+  "display_hint": null,
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 1 quarter-turn(s); gravity-verified at export)",
   "rgb_native_resolution_hw": [
     1920,
     1440
@@ -32,6 +36,7 @@
       1.0
     ]
   ],
+  "intrinsics_frame": "native raster",
   "depth_intrinsics_3x3": [
     [
       197.02384440104166,

episodes/019fb6df-0a15-7630-b4e1-b5e552791fcc/openego_sidecar.json

@@ -422,79 +422,100 @@
       },
       "quality_gates": {
         "all_passed": true,
+        "validated_at": "2026-08-10T13:34:24+00:00",
         "gates": [
           {
-            "detail": "",
             "name": "json_schema",
             "passed": true,
+            "value": "ok",
             "threshold": "validates",
-            "value": "ok"
+            "detail": ""
           },
           {
-            "detail": "presence-conditioned over the actor's active span [126..719] = 594/778 frames (raw all-frames = 31.36%)",
             "name": "frames_with_hands_pct",
             "passed": true,
-            "threshold": ">= 30% [class=GLOBAL]",
-            "value": 41.08
+            "value": 100.0,
+            "threshold": ">= 5% [class=sweep]",
+            "detail": "presence-conditioned over the actor's active span [47..727] = 681/778 frames (raw all-frames = 87.53%); shipped channel = raw \u2295 hand_mask_corrections"
           },
           {
-            "detail": "silent or near-silent",
             "name": "audio_snr_db",
             "passed": true,
+            "value": null,
             "threshold": ">= 5 dB",
-            "value": null
+            "detail": "silent or near-silent"
           },
           {
-            "detail": "silent clip",
             "name": "transcript_density",
             "passed": true,
+            "value": 0.0,
             "threshold": ">= 0.5 wps",
-            "value": 0.0
+            "detail": "silent clip"
           },
           {
-            "detail": "",
             "name": "hand_mask_coverage_pct",
             "passed": true,
+            "value": 1.917,
             "threshold": ">= 1%",
-            "value": 1.917
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "phase_plan_windowing",
             "passed": true,
+            "value": "not_requested",
             "threshold": "phase plan absent \u2014 skipped",
-            "value": "not_requested"
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "object_mask_coverage_pct",
             "passed": true,
-            "threshold": ">= 1% (per present object) [class=GLOBAL]",
-            "value": 3.923
+            "value": 3.923,
+            "threshold": ">= 0.5% (per present object) [class=sweep]",
+            "detail": ""
           },
           {
-            "detail": "left hand absent \u2014 skipped",
             "name": "hand_mask_area_floor_left",
             "passed": true,
+            "value": null,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": null
+            "detail": "left hand absent \u2014 skipped"
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_right",
             "passed": true,
+            "value": 4.1835,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 4.1835
+            "detail": ""
           },
           {
-            "detail": "both hands not present \u2014 skipped",
             "name": "hand_mask_lr_asymmetry",
             "passed": true,
+            "value": null,
             "threshold": "min/max >= 0.20",
-            "value": null
+            "detail": "both hands not present \u2014 skipped"
+          },
+          {
+            "name": "hand_slot_merge",
+            "passed": true,
+            "value": 0,
+            "threshold": "< 5 consecutive frames with L\u2229R IoU > 0.5",
+            "detail": ""
+          },
+          {
+            "name": "object_track_truncated",
+            "passed": true,
+            "value": 0,
+            "threshold": "no tracked object ends while a hand still holds it",
+            "detail": ""
+          },
+          {
+            "name": "actor_mirror_consistency",
+            "passed": true,
+            "value": 0,
+            "threshold": "every hierarchy span's actor_kind == its action's actors[0]",
+            "detail": ""
           }
-        ],
-        "validated_at": "2026-07-31T08:26:03+00:00"
+        ]
       }
     },
     "manually_annotated_fields": [],

episodes/019fb7cd-b333-7b73-a360-149207d553c6/depth_geometry.json

@@ -0,0 +1,62 @@
+{
+  "sensor": "iphone_lidar_arkit",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "depth_resolution_hw": [
+    256,
+    192
+  ],
+  "ego_obs_resolution_hw": [
+    512,
+    384
+  ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": "ccw90",
+  "display_hint": "rotate 90\u00b0 CW to view upright",
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 0 quarter-turn(s); gravity-verified at export)",
+  "rgb_native_resolution_hw": [
+    1440,
+    1920
+  ],
+  "depth_output_unit_default": "mm (LeRobot depth_output_unit default; pass --dataset.depth_output_unit m to load metres)",
+  "raw_depth_reference": "openego_sidecar.json + data/annotations/<cid>/depth.npz (lossless)",
+  "arkit_rgb_intrinsics_3x3": [
+    [
+      1471.7620849609375,
+      0.0,
+      745.2727661132812
+    ],
+    [
+      0.0,
+      1471.7620849609375,
+      947.3223266601562
+    ],
+    [
+      0.0,
+      0.0,
+      1.0
+    ]
+  ],
+  "intrinsics_frame": "shipped raster (post-rotation)",
+  "depth_intrinsics_3x3": [
+    [
+      196.23494466145834,
+      0.0,
+      99.3697021484375
+    ],
+    [
+      0.0,
+      196.23494466145834,
+      126.3096435546875
+    ],
+    [
+      0.0,
+      0.0,
+      1.0
+    ]
+  ],
+  "depth_intrinsics_scale_xy": [
+    0.13333333333333333,
+    0.13333333333333333
+  ],
+  "intrinsics_source": "record3d_arkit (median of per-frame, via chokepoint)"
+}

episodes/{019fb7c0-1159-7242-9c36-bcac261ec817 → 019fb7cd-b333-7b73-a360-149207d553c6}/hand_object_contact.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5ec3ac6ee043cf13340d7304d16d1d2d92f12eb5740d929d9a77c95f1f94d179
-size 19092
+oid sha256:ed9021cc8960ff1082e1758682f0405f5cc90c648d4d886a54223c3619abe8bd
+size 11058

episodes/019fb7cd-b333-7b73-a360-149207d553c6/language_spans.json

@@ -0,0 +1,73 @@
+{
+  "task": "sweep",
+  "n_spans": 3,
+  "n_keysteps": 3,
+  "spans": [
+    {
+      "start_s": 0.0,
+      "end_s": 3.0,
+      "label": "grasp dustpan on towel",
+      "objects": [
+        "dustpan"
+      ],
+      "actors": [
+        "right_hand"
+      ],
+      "source": "human",
+      "confidence": "low",
+      "grounded": false,
+      "evidence": "The dustpan rests on the striped towel; the right hand takes hold of it. The brush has not yet entered frame."
+    },
+    {
+      "start_s": 3.0,
+      "end_s": 4.5,
+      "label": "bring brush to dustpan",
+      "objects": [
+        "brush"
+      ],
+      "actors": [
+        "both_hands"
+      ],
+      "source": "human",
+      "confidence": "high",
+      "grounded": true,
+      "evidence": "The brush is brought to the dustpan positioned on the towel."
+    },
+    {
+      "start_s": 4.5,
+      "end_s": 13.333,
+      "label": "sweep debris into dustpan",
+      "objects": [
+        "debris"
+      ],
+      "actors": [
+        "both_hands"
+      ],
+      "source": "human",
+      "confidence": "high",
+      "grounded": true,
+      "evidence": "Brush pushes crumpled white paper into the white dustpan."
+    }
+  ],
+  "keysteps": [
+    {
+      "start_s": 0.0,
+      "end_s": 3.0,
+      "description": "position dustpan on towel",
+      "essential": true
+    },
+    {
+      "start_s": 3.0,
+      "end_s": 4.5,
+      "description": "bring brush into frame",
+      "essential": true
+    },
+    {
+      "start_s": 4.5,
+      "end_s": 13.333,
+      "description": "sweep paper into dustpan",
+      "essential": true
+    }
+  ],
+  "note": "Time-aligned sub-step language \u2014 the VLM-partitioned action `spans` (label + objects + actors + VLM evidence/confidence/grounded) and the coarser `keysteps`. This is a RIDER, not per-frame: the LeRobot `task` column stays the static high-level goal so it never flips frame-to-frame; use these spans for language-conditioned / sub-goal training by joining on the `timestamp` (seconds). A span with `grounded=false` or a low `confidence` is an honest low-trust label \u2014 filter accordingly."
+}

episodes/{019fb7c6-01ed-7ca3-aa35-11b87a8a8402 → 019fb7cd-b333-7b73-a360-149207d553c6}/object_mask.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5f93cc2355b15de1f6bbf43e228b5a077b21914add4424d46481e337e120bc35
-size 1202066
+oid sha256:88ddfed5e1602225fa32c0bee6f624220c0539b5a913ece99b715d424e095eab
+size 3904829

episodes/019fb7cd-b333-7b73-a360-149207d553c6/object_object_contact.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90a66643f9a46f7798c9f9ff9a07959485e2cd49f83d174a8c61cfcc2fe6a251
+size 3373

episodes/019fb7cd-b333-7b73-a360-149207d553c6/object_registry.json

@@ -0,0 +1,17 @@
+{
+  "version": "0.1.0",
+  "objects": [
+    {
+      "object_id": "obj_brush",
+      "label": "small brush",
+      "phrase": "brush",
+      "color_idx": 1
+    },
+    {
+      "object_id": "obj_dustpan",
+      "label": "dustpan",
+      "phrase": "dustpan",
+      "color_idx": 3
+    }
+  ]
+}

episodes/019fbc13-d77f-70c1-b142-4b07a7df35f6/contact_corrections.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edff5c61300a264ac2e9ac681e96e4e368eec4cb51079022230937d61ea01191
+size 2173

episodes/019fbc13-d77f-70c1-b142-4b07a7df35f6/depth_geometry.json

@@ -0,0 +1,62 @@
+{
+  "sensor": "iphone_lidar_arkit",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "depth_resolution_hw": [
+    256,
+    192
+  ],
+  "ego_obs_resolution_hw": [
+    512,
+    384
+  ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": "ccw90",
+  "display_hint": "rotate 90\u00b0 CW to view upright",
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 0 quarter-turn(s); gravity-verified at export)",
+  "rgb_native_resolution_hw": [
+    1440,
+    1920
+  ],
+  "depth_output_unit_default": "mm (LeRobot depth_output_unit default; pass --dataset.depth_output_unit m to load metres)",
+  "raw_depth_reference": "openego_sidecar.json + data/annotations/<cid>/depth.npz (lossless)",
+  "arkit_rgb_intrinsics_3x3": [
+    [
+      1472.691162109375,
+      0.0,
+      742.8125
+    ],
+    [
+      0.0,
+      1472.691162109375,
+      953.4917602539062
+    ],
+    [
+      0.0,
+      0.0,
+      1.0
+    ]
+  ],
+  "intrinsics_frame": "shipped raster (post-rotation)",
+  "depth_intrinsics_3x3": [
+    [
+      196.35882161458332,
+      0.0,
+      99.04166666666667
+    ],
+    [
+      0.0,
+      196.35882161458332,
+      127.13223470052083
+    ],
+    [
+      0.0,
+      0.0,
+      1.0
+    ]
+  ],
+  "depth_intrinsics_scale_xy": [
+    0.13333333333333333,
+    0.13333333333333333
+  ],
+  "intrinsics_source": "record3d_arkit (median of per-frame, via chokepoint)"
+}

episodes/{019fb6db-ff4d-7f00-a9ae-58c76698a120 → 019fbc13-d77f-70c1-b142-4b07a7df35f6}/hand_object_contact.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:204f6ab45205164bf78d5ce2b7aab498e65c96718dad3b3f4cd70f8c939f9955
-size 15137
+oid sha256:dc06dcdb53553eafe8a9b4159b569321a5820df568bc61b1304c6f57d8ce43a4
+size 13659

episodes/019fbc13-d77f-70c1-b142-4b07a7df35f6/language_spans.json

@@ -0,0 +1,115 @@
+{
+  "task": "scoop_grain",
+  "n_spans": 5,
+  "n_keysteps": 5,
+  "spans": [
+    {
+      "start_s": 0.0,
+      "end_s": 5.5,
+      "label": "pick up mug",
+      "objects": [
+        "mug"
+      ],
+      "actors": [
+        "right_hand"
+      ],
+      "source": "vlm",
+      "confidence": "high",
+      "grounded": true,
+      "evidence": "Hands approach and grasp the mug, lifting it slightly from the table."
+    },
+    {
+      "start_s": 5.5,
+      "end_s": 9.0,
+      "label": "move mug to bowl",
+      "objects": [
+        "mug"
+      ],
+      "actors": [
+        "right_hand"
+      ],
+      "source": "vlm",
+      "confidence": "high",
+      "grounded": true,
+      "evidence": "The mug is lifted from the table and moved over the large bowl containing grains."
+    },
+    {
+      "start_s": 9.0,
+      "end_s": 12.0,
+      "label": "scoop grains from bowl",
+      "objects": [
+        "grains"
+      ],
+      "actors": [
+        "right_hand"
+      ],
+      "source": "vlm",
+      "confidence": "low",
+      "grounded": false,
+      "evidence": "The mug is dipped into the large bowl, filled with grains, and then lifted."
+    },
+    {
+      "start_s": 12.0,
+      "end_s": 17.0,
+      "label": "pour grains into saucepan",
+      "objects": [
+        "grains"
+      ],
+      "actors": [
+        "right_hand"
+      ],
+      "source": "vlm",
+      "confidence": "low",
+      "grounded": false,
+      "evidence": "The mug is tilted over the saucepan and grains fall from it into the pan."
+    },
+    {
+      "start_s": 17.0,
+      "end_s": 22.933,
+      "label": "place mug on table",
+      "objects": [
+        "grains"
+      ],
+      "actors": [
+        "both_hands"
+      ],
+      "source": "vlm",
+      "confidence": "low",
+      "grounded": false,
+      "evidence": "The emptied mug is lowered and set down on the table."
+    }
+  ],
+  "keysteps": [
+    {
+      "start_s": 0.0,
+      "end_s": 5.5,
+      "description": "pick up mug",
+      "essential": true
+    },
+    {
+      "start_s": 5.5,
+      "end_s": 9.0,
+      "description": "move mug to bowl",
+      "essential": true
+    },
+    {
+      "start_s": 9.0,
+      "end_s": 12.0,
+      "description": "scoop grains from bowl",
+      "essential": true
+    },
+    {
+      "start_s": 12.0,
+      "end_s": 17.0,
+      "description": "pour grains into saucepan",
+      "essential": true
+    },
+    {
+      "start_s": 17.0,
+      "end_s": 22.933,
+      "description": "place mug on table",
+      "essential": true
+    }
+  ],
+  "note": "Time-aligned sub-step language \u2014 the VLM-partitioned action `spans` (label + objects + actors + VLM evidence/confidence/grounded) and the coarser `keysteps`. This is a RIDER, not per-frame: the LeRobot `task` column stays the static high-level goal so it never flips frame-to-frame; use these spans for language-conditioned / sub-goal training by joining on the `timestamp` (seconds). A span with `grounded=false` or a low `confidence` is an honest low-trust label \u2014 filter accordingly."
+}

episodes/{019fbc1c-81a7-7921-9bd2-aa8973f2959b → 019fbc13-d77f-70c1-b142-4b07a7df35f6}/object_mask.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3fe7f8ee752937f1b3600b4ada883c4334954d2958899029ab58cb9a62d26c4d
-size 1288445
+oid sha256:793bc43e6594ffd0fb8ec112231f50f9d4491348f506777f2945409d2a502e0e
+size 3803418

episodes/019fbc15-11a7-7860-889f-d07cfa430507/depth_geometry.json

@@ -1,14 +1,18 @@
 {
   "sensor": "iphone_lidar_arkit",
-  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth, its OWN intrinsics \u2014 see depth_geometry.json); observation.images.ego is 256x256 (resized). DIFFERENT pixel grids \u2014 register via the depth intrinsics + observation.camera_pose, never depth[y,x]==ego[y,x]. The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
   "depth_resolution_hw": [
     256,
     192
   ],
   "ego_obs_resolution_hw": [
-    256,
-    256
+    512,
+    384
   ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": null,
+  "display_hint": null,
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 1 quarter-turn(s); gravity-verified at export)",
   "rgb_native_resolution_hw": [
     1920,
     1440
@@ -32,6 +36,7 @@
       1.0
     ]
   ],
+  "intrinsics_frame": "native raster",
   "depth_intrinsics_3x3": [
     [
       196.59479166666665,

episodes/019fbc15-11a7-7860-889f-d07cfa430507/openego_sidecar.json

@@ -399,86 +399,100 @@
       },
       "quality_gates": {
         "all_passed": true,
+        "validated_at": "2026-08-10T13:30:55+00:00",
         "gates": [
           {
-            "detail": "",
             "name": "json_schema",
             "passed": true,
+            "value": "ok",
             "threshold": "validates",
-            "value": "ok"
+            "detail": ""
           },
           {
-            "detail": "presence-conditioned over the actor's active span [92..634] = 543/676 frames (raw all-frames = 79.88%)",
             "name": "frames_with_hands_pct",
             "passed": true,
-            "threshold": ">= 30% [class=GLOBAL]",
-            "value": 99.45
+            "value": 100.0,
+            "threshold": ">= 5% [class=sweep]",
+            "detail": "presence-conditioned over the actor's active span [86..643] = 558/676 frames (raw all-frames = 82.54%); shipped channel = raw \u2295 hand_mask_corrections"
           },
           {
-            "detail": "silent or near-silent",
             "name": "audio_snr_db",
             "passed": true,
+            "value": null,
             "threshold": ">= 5 dB",
-            "value": null
+            "detail": "silent or near-silent"
           },
           {
-            "detail": "silent clip",
             "name": "transcript_density",
             "passed": true,
+            "value": 0.0,
             "threshold": ">= 0.5 wps",
-            "value": 0.0
+            "detail": "silent clip"
           },
           {
-            "detail": "",
             "name": "hand_mask_coverage_pct",
             "passed": true,
+            "value": 3.143,
             "threshold": ">= 1%",
-            "value": 3.143
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "phase_plan_windowing",
             "passed": true,
+            "value": "not_requested",
             "threshold": "phase plan absent \u2014 skipped",
-            "value": "not_requested"
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "object_mask_coverage_pct",
             "passed": true,
-            "threshold": ">= 1% (per present object) [class=GLOBAL]",
-            "value": 3.019
+            "value": 3.019,
+            "threshold": ">= 0.5% (per present object) [class=sweep]",
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_left",
             "passed": true,
+            "value": 3.2262,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 3.2262
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_right",
             "passed": true,
+            "value": 4.7846,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 4.7846
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_lr_asymmetry",
             "passed": true,
+            "value": 0.6743,
             "threshold": "min/max >= 0.20",
-            "value": 0.6743
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_slot_merge",
             "passed": true,
+            "value": 0,
             "threshold": "< 5 consecutive frames with L\u2229R IoU > 0.5",
-            "value": 0
+            "detail": ""
+          },
+          {
+            "name": "object_track_truncated",
+            "passed": true,
+            "value": 0,
+            "threshold": "no tracked object ends while a hand still holds it",
+            "detail": ""
+          },
+          {
+            "name": "actor_mirror_consistency",
+            "passed": true,
+            "value": 0,
+            "threshold": "every hierarchy span's actor_kind == its action's actors[0]",
+            "detail": ""
           }
-        ],
-        "validated_at": "2026-08-01T15:22:41+00:00"
+        ]
       }
     },
     "manually_annotated_fields": [],

episodes/019fbc16-b84b-7ad3-8387-7a3305b151d0/depth_geometry.json

@@ -1,14 +1,18 @@
 {
   "sensor": "iphone_lidar_arkit",
-  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth, its OWN intrinsics \u2014 see depth_geometry.json); observation.images.ego is 256x256 (resized). DIFFERENT pixel grids \u2014 register via the depth intrinsics + observation.camera_pose, never depth[y,x]==ego[y,x]. The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
   "depth_resolution_hw": [
     256,
     192
   ],
   "ego_obs_resolution_hw": [
-    256,
-    256
+    512,
+    384
   ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": null,
+  "display_hint": null,
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 1 quarter-turn(s); gravity-verified at export)",
   "rgb_native_resolution_hw": [
     1920,
     1440
@@ -32,6 +36,7 @@
       1.0
     ]
   ],
+  "intrinsics_frame": "native raster",
   "depth_intrinsics_3x3": [
     [
       196.544970703125,

episodes/019fbc16-b84b-7ad3-8387-7a3305b151d0/openego_sidecar.json

@@ -451,86 +451,100 @@
       },
       "quality_gates": {
         "all_passed": true,
+        "validated_at": "2026-08-10T13:30:42+00:00",
         "gates": [
           {
-            "detail": "",
             "name": "json_schema",
             "passed": true,
+            "value": "ok",
             "threshold": "validates",
-            "value": "ok"
+            "detail": ""
           },
           {
-            "detail": "presence-conditioned over the actor's active span [98..685] = 588/757 frames (raw all-frames = 71.07%)",
             "name": "frames_with_hands_pct",
             "passed": true,
-            "threshold": ">= 30% [class=GLOBAL]",
-            "value": 91.5
+            "value": 100.0,
+            "threshold": ">= 5% [class=sweep]",
+            "detail": "presence-conditioned over the actor's active span [92..698] = 607/757 frames (raw all-frames = 80.18%); shipped channel = raw \u2295 hand_mask_corrections"
           },
           {
-            "detail": "silent or near-silent",
             "name": "audio_snr_db",
             "passed": true,
+            "value": null,
             "threshold": ">= 5 dB",
-            "value": null
+            "detail": "silent or near-silent"
           },
           {
-            "detail": "silent clip",
             "name": "transcript_density",
             "passed": true,
+            "value": 0.0,
             "threshold": ">= 0.5 wps",
-            "value": 0.0
+            "detail": "silent clip"
           },
           {
-            "detail": "",
             "name": "hand_mask_coverage_pct",
             "passed": true,
+            "value": 2.675,
             "threshold": ">= 1%",
-            "value": 2.675
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "phase_plan_windowing",
             "passed": true,
+            "value": "not_requested",
             "threshold": "phase plan absent \u2014 skipped",
-            "value": "not_requested"
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "object_mask_coverage_pct",
             "passed": true,
-            "threshold": ">= 1% (per present object) [class=GLOBAL]",
-            "value": 3.594
+            "value": 3.594,
+            "threshold": ">= 0.5% (per present object) [class=sweep]",
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_left",
             "passed": true,
+            "value": 2.702,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 2.702
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_right",
             "passed": true,
+            "value": 3.774,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 3.774
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_lr_asymmetry",
             "passed": true,
+            "value": 0.716,
             "threshold": "min/max >= 0.20",
-            "value": 0.716
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_slot_merge",
             "passed": true,
+            "value": 0,
             "threshold": "< 5 consecutive frames with L\u2229R IoU > 0.5",
-            "value": 0
+            "detail": ""
+          },
+          {
+            "name": "object_track_truncated",
+            "passed": true,
+            "value": 0,
+            "threshold": "no tracked object ends while a hand still holds it",
+            "detail": ""
+          },
+          {
+            "name": "actor_mirror_consistency",
+            "passed": true,
+            "value": 0,
+            "threshold": "every hierarchy span's actor_kind == its action's actors[0]",
+            "detail": ""
           }
-        ],
-        "validated_at": "2026-08-01T17:15:36+00:00"
+        ]
       }
     },
     "manually_annotated_fields": [],

episodes/019fbc18-33ed-7270-8166-1997f31222c8/depth_geometry.json

@@ -1,14 +1,18 @@
 {
   "sensor": "iphone_lidar_arkit",
-  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth, its OWN intrinsics \u2014 see depth_geometry.json); observation.images.ego is 256x256 (resized). DIFFERENT pixel grids \u2014 register via the depth intrinsics + observation.camera_pose, never depth[y,x]==ego[y,x]. The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
   "depth_resolution_hw": [
     256,
     192
   ],
   "ego_obs_resolution_hw": [
-    256,
-    256
+    512,
+    384
   ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": null,
+  "display_hint": null,
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 1 quarter-turn(s); gravity-verified at export)",
   "rgb_native_resolution_hw": [
     1920,
     1440
@@ -32,6 +36,7 @@
       1.0
     ]
   ],
+  "intrinsics_frame": "native raster",
   "depth_intrinsics_3x3": [
     [
       196.660888671875,

episodes/019fbc18-33ed-7270-8166-1997f31222c8/openego_sidecar.json

@@ -522,7 +522,7 @@
       },
       "quality_gates": {
         "all_passed": true,
-        "validated_at": "2026-08-02T07:59:32+00:00",
+        "validated_at": "2026-08-10T13:30:28+00:00",
         "gates": [
           {
             "name": "json_schema",
@@ -534,9 +534,9 @@
           {
             "name": "frames_with_hands_pct",
             "passed": true,
-            "value": 75.99,
-            "threshold": ">= 30% [class=GLOBAL]",
-            "detail": "presence-conditioned over the actor's active span [97..654] = 558/698 frames (raw all-frames = 60.74%)"
+            "value": 100.0,
+            "threshold": ">= 5% [class=sweep]",
+            "detail": "presence-conditioned over the actor's active span [98..661] = 564/698 frames (raw all-frames = 80.80%); shipped channel = raw \u2295 hand_mask_corrections"
           },
           {
             "name": "audio_snr_db",
@@ -570,7 +570,7 @@
             "name": "object_mask_coverage_pct",
             "passed": true,
             "value": 3.88,
-            "threshold": ">= 1% (per present object) [class=GLOBAL]",
+            "threshold": ">= 0.5% (per present object) [class=sweep]",
             "detail": ""
           },
           {
@@ -600,6 +600,20 @@
             "value": 0,
             "threshold": "< 5 consecutive frames with L\u2229R IoU > 0.5",
             "detail": ""
+          },
+          {
+            "name": "object_track_truncated",
+            "passed": true,
+            "value": 0,
+            "threshold": "no tracked object ends while a hand still holds it",
+            "detail": ""
+          },
+          {
+            "name": "actor_mirror_consistency",
+            "passed": true,
+            "value": 0,
+            "threshold": "every hierarchy span's actor_kind == its action's actors[0]",
+            "detail": ""
           }
         ]
       }

episodes/019fbc1c-81a7-7921-9bd2-aa8973f2959b/hand_object_contact.npz

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:13ba6e02858624f8c0319bdb5c1dfe7011605461d6a375ced6d3b2f257f2bb21
-size 150588

episodes/019fbdf1-dadb-7400-bca3-1382d9d9c666/depth_geometry.json

@@ -1,14 +1,18 @@
 {
   "sensor": "iphone_lidar_arkit",
-  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth, its OWN intrinsics \u2014 see depth_geometry.json); observation.images.ego is 256x256 (resized). DIFFERENT pixel grids \u2014 register via the depth intrinsics + observation.camera_pose, never depth[y,x]==ego[y,x]. The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
   "depth_resolution_hw": [
     256,
     192
   ],
   "ego_obs_resolution_hw": [
-    256,
-    256
+    512,
+    384
   ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": null,
+  "display_hint": null,
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 1 quarter-turn(s); gravity-verified at export)",
   "rgb_native_resolution_hw": [
     1920,
     1440
@@ -32,6 +36,7 @@
       1.0
     ]
   ],
+  "intrinsics_frame": "native raster",
   "depth_intrinsics_3x3": [
     [
       183.91941731770834,

episodes/019fbdf1-dadb-7400-bca3-1382d9d9c666/language_spans.json

@@ -1,19 +1,19 @@
 {
   "task": "fold_laundry",
-  "n_spans": 5,
-  "n_keysteps": 5,
+  "n_spans": 4,
+  "n_keysteps": 4,
   "spans": [
     {
       "start_s": 0.0,
       "end_s": 2.5,
-      "label": "arrange shorts on surface",
+      "label": "stretch shorts out",
       "objects": [
         "shorts"
       ],
       "actors": [
         "both_hands"
       ],
-      "source": "vlm",
+      "source": "human",
       "confidence": "high",
       "grounded": true,
       "evidence": "Hands enter frame and flatten the shorts on the quilted surface."
@@ -36,47 +36,32 @@
     {
       "start_s": 7.5,
       "end_s": 11.5,
-      "label": "fold shorts horizontally",
+      "label": "fold shorts widthwise",
       "objects": [
         "shorts"
       ],
       "actors": [
         "both_hands"
       ],
-      "source": "vlm",
+      "source": "human",
       "confidence": "high",
       "grounded": true,
-      "evidence": "Hands fold the shorts again along the horizontal axis, forming a compact shape."
+      "evidence": "Hands fold the shorts in half widthwise, forming a compact bundle."
     },
     {
       "start_s": 11.5,
-      "end_s": 12.3,
-      "label": "press shorts into shape",
-      "objects": [
-        "shorts"
-      ],
-      "actors": [
-        "both_hands"
-      ],
-      "source": "vlm",
-      "confidence": "high",
-      "grounded": true,
-      "evidence": "Hands press down on the rolled shorts to compact them."
-    },
-    {
-      "start_s": 12.3,
       "end_s": 14.067,
-      "label": "leave folded shorts on surface",
+      "label": "flatten shorts on bed",
       "objects": [
         "shorts"
       ],
       "actors": [
         "both_hands"
       ],
-      "source": "vlm",
+      "source": "human",
       "confidence": "high",
       "grounded": true,
-      "evidence": "Hands move out of frame, leaving the compactly folded shorts."
+      "evidence": "Hands press down on the folded shorts to compact them on the bed."
     }
   ],
   "keysteps": [
@@ -100,14 +85,8 @@
     },
     {
       "start_s": 11.5,
-      "end_s": 12.3,
-      "description": "press shorts into shape",
-      "essential": true
-    },
-    {
-      "start_s": 12.3,
       "end_s": 14.067,
-      "description": "leave folded shorts on surface",
+      "description": "press shorts into shape",
       "essential": true
     }
   ],

episodes/019fbdf1-dadb-7400-bca3-1382d9d9c666/openego_sidecar.json

@@ -8,7 +8,8 @@
       "extensions": {
         "actor_source": "motion",
         "object_source": "vlm",
-        "span_source": "vlm",
+        "span_id": "s0",
+        "span_source": "human",
         "vlm_confidence": "high",
         "vlm_evidence": "Hands enter frame and flatten the shorts on the quilted surface.",
         "vlm_grounded": true,
@@ -27,7 +28,7 @@
           }
         ]
       },
-      "label": "arrange shorts on surface",
+      "label": "stretch shorts out",
       "narration_uncertain": false,
       "objects": [
         "shorts"
@@ -42,6 +43,7 @@
       "extensions": {
         "actor_source": "contact",
         "object_source": "vlm",
+        "span_id": "s1",
         "span_source": "vlm",
         "vlm_confidence": "high",
         "vlm_evidence": "Hands fold the shorts in half along their length.",
@@ -76,9 +78,10 @@
       "extensions": {
         "actor_source": "contact",
         "object_source": "vlm",
-        "span_source": "vlm",
+        "span_id": "s2",
+        "span_source": "human",
         "vlm_confidence": "high",
-        "vlm_evidence": "Hands fold the shorts again along the horizontal axis, forming a compact shape.",
+        "vlm_evidence": "Hands fold the shorts in half widthwise, forming a compact bundle.",
         "vlm_grounded": true,
         "vlm_object_candidates": [
           {
@@ -95,7 +98,7 @@
           }
         ]
       },
-      "label": "fold shorts horizontally",
+      "label": "fold shorts widthwise",
       "narration_uncertain": false,
       "objects": [
         "shorts"
@@ -106,13 +109,14 @@
       "actors": [
         "both_hands"
       ],
-      "end_timestamp": 12.3,
+      "end_timestamp": 14.067,
       "extensions": {
         "actor_source": "contact",
         "object_source": "vlm",
-        "span_source": "vlm",
+        "span_id": "s3",
+        "span_source": "human",
         "vlm_confidence": "high",
-        "vlm_evidence": "Hands press down on the rolled shorts to compact them.",
+        "vlm_evidence": "Hands press down on the folded shorts to compact them on the bed.",
         "vlm_grounded": true,
         "vlm_object_candidates": [
           {
@@ -129,40 +133,12 @@
           }
         ]
       },
-      "label": "press shorts into shape",
+      "label": "flatten shorts on bed",
       "narration_uncertain": false,
       "objects": [
         "shorts"
       ],
       "start_timestamp": 11.5
-    },
-    {
-      "actors": [
-        "both_hands"
-      ],
-      "end_timestamp": 14.067,
-      "extensions": {
-        "actor_source": "contact",
-        "object_source": "vlm",
-        "span_source": "vlm",
-        "vlm_confidence": "high",
-        "vlm_evidence": "Hands move out of frame, leaving the compactly folded shorts.",
-        "vlm_grounded": true,
-        "vlm_object_candidates": [
-          {
-            "evidence": "compactly folded shorts remaining on the surface",
-            "label": "shorts",
-            "role": "workpiece",
-            "track": true
-          }
-        ]
-      },
-      "label": "leave folded shorts on surface",
-      "narration_uncertain": false,
-      "objects": [
-        "shorts"
-      ],
-      "start_timestamp": 12.3
     }
   ],
   "annotation_metadata": {
@@ -414,100 +390,100 @@
       },
       "quality_gates": {
         "all_passed": true,
+        "validated_at": "2026-08-10T13:36:05+00:00",
         "gates": [
           {
-            "detail": "",
             "name": "json_schema",
             "passed": true,
+            "value": "ok",
             "threshold": "validates",
-            "value": "ok"
+            "detail": ""
           },
           {
-            "detail": "presence-conditioned over the actor's active span [105..755] = 651/844 frames (raw all-frames = 72.51%)",
             "name": "frames_with_hands_pct",
             "passed": true,
+            "value": 100.0,
             "threshold": ">= 30% [class=GLOBAL]",
-            "value": 94.01
+            "detail": "presence-conditioned over the actor's active span [100..759] = 660/844 frames (raw all-frames = 78.20%); shipped channel = raw \u2295 hand_mask_corrections"
           },
           {
-            "detail": "silent or near-silent",
             "name": "audio_snr_db",
             "passed": true,
+            "value": null,
             "threshold": ">= 5 dB",
-            "value": null
+            "detail": "silent or near-silent"
           },
           {
-            "detail": "silent clip",
             "name": "transcript_density",
             "passed": true,
+            "value": 0.0,
             "threshold": ">= 0.5 wps",
-            "value": 0.0
+            "detail": "silent clip"
           },
           {
-            "detail": "",
             "name": "hand_mask_coverage_pct",
             "passed": true,
+            "value": 3.337,
             "threshold": ">= 1%",
-            "value": 3.337
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "phase_plan_windowing",
             "passed": true,
+            "value": "not_requested",
             "threshold": "phase plan absent \u2014 skipped",
-            "value": "not_requested"
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "object_mask_coverage_pct",
             "passed": true,
+            "value": 27.009,
             "threshold": ">= 1% (per present object) [class=GLOBAL]",
-            "value": 27.009
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_left",
             "passed": true,
+            "value": 3.7305,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 3.7305
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_right",
             "passed": true,
+            "value": 4.4407,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 4.4407
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_lr_asymmetry",
             "passed": true,
+            "value": 0.8401,
             "threshold": "min/max >= 0.20",
-            "value": 0.8401
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_slot_merge",
             "passed": true,
+            "value": 0,
             "threshold": "< 5 consecutive frames with L\u2229R IoU > 0.5",
-            "value": 0
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "object_track_truncated",
             "passed": true,
+            "value": 0,
             "threshold": "no tracked object ends while a hand still holds it",
-            "value": 0
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "actor_mirror_consistency",
             "passed": true,
+            "value": 0,
             "threshold": "every hierarchy span's actor_kind == its action's actors[0]",
-            "value": 0
+            "detail": ""
           }
-        ],
-        "validated_at": "2026-08-04T09:08:10+00:00"
+        ]
       }
     },
     "manually_annotated_fields": [],
@@ -19633,7 +19609,7 @@
           "end_time": 2.5,
           "is_essential": true,
           "keystep_ref": 0,
-          "label": "arrange shorts on surface",
+          "label": "stretch shorts out",
           "node_id": "e0",
           "spans": [
             {
@@ -19643,7 +19619,7 @@
               "end_time": 2.5,
               "evidence": "Hands enter frame and flatten the shorts on the quilted surface.",
               "grounded": true,
-              "label": "arrange shorts on surface",
+              "label": "stretch shorts out",
               "object": "shorts",
               "object_candidates": [
                 {
@@ -19660,6 +19636,7 @@
                 }
               ],
               "span_id": "s0",
+              "span_source": "human",
               "start_time": 0.0
             }
           ],
@@ -19707,7 +19684,7 @@
           "end_time": 11.5,
           "is_essential": true,
           "keystep_ref": 2,
-          "label": "fold shorts horizontally",
+          "label": "fold shorts widthwise",
           "node_id": "e2",
           "spans": [
             {
@@ -19715,9 +19692,9 @@
               "actor_kind": "both_hands",
               "confidence": "high",
               "end_time": 11.5,
-              "evidence": "Hands fold the shorts again along the horizontal axis, forming a compact shape.",
+              "evidence": "Hands fold the shorts in half widthwise, forming a compact bundle.",
               "grounded": true,
-              "label": "fold shorts horizontally",
+              "label": "fold shorts widthwise",
               "object": "shorts",
               "object_candidates": [
                 {
@@ -19734,6 +19711,7 @@
                 }
               ],
               "span_id": "s2",
+              "span_source": "human",
               "start_time": 7.5
             }
           ],
@@ -19741,20 +19719,20 @@
           "sub_goal": "fold shorts"
         },
         {
-          "end_time": 12.3,
+          "end_time": 14.067,
           "is_essential": true,
           "keystep_ref": 3,
-          "label": "press shorts into shape",
+          "label": "flatten shorts on bed",
           "node_id": "e3",
           "spans": [
             {
               "action_ref": 3,
               "actor_kind": "both_hands",
               "confidence": "high",
-              "end_time": 12.3,
-              "evidence": "Hands press down on the rolled shorts to compact them.",
+              "end_time": 14.067,
+              "evidence": "Hands press down on the folded shorts to compact them on the bed.",
               "grounded": true,
-              "label": "press shorts into shape",
+              "label": "flatten shorts on bed",
               "object": "shorts",
               "object_candidates": [
                 {
@@ -19771,42 +19749,12 @@
                 }
               ],
               "span_id": "s3",
+              "span_source": "human",
               "start_time": 11.5
             }
           ],
           "start_time": 11.5,
           "sub_goal": "fold shorts"
-        },
-        {
-          "end_time": 14.067,
-          "is_essential": true,
-          "keystep_ref": 4,
-          "label": "leave folded shorts on surface",
-          "node_id": "e4",
-          "spans": [
-            {
-              "action_ref": 4,
-              "actor_kind": "both_hands",
-              "confidence": "high",
-              "end_time": 14.067,
-              "evidence": "Hands move out of frame, leaving the compactly folded shorts.",
-              "grounded": true,
-              "label": "leave folded shorts on surface",
-              "object": "shorts",
-              "object_candidates": [
-                {
-                  "evidence": "compactly folded shorts remaining on the surface",
-                  "label": "shorts",
-                  "role": "workpiece",
-                  "track": true
-                }
-              ],
-              "span_id": "s4",
-              "start_time": 12.3
-            }
-          ],
-          "start_time": 12.3,
-          "sub_goal": "fold shorts"
         }
       ],
       "span_source": "vlm",
@@ -32561,8 +32509,8 @@
       "start_time": 0.0,
       "step_description": "arrange shorts on surface",
       "step_id": 0,
-      "step_name": "arrange shorts on surface",
-      "step_unique_id": 103048908367292
+      "step_name": "stretch shorts out",
+      "step_unique_id": 2191387116819
     },
     {
       "best_view": "ego",
@@ -32583,30 +32531,19 @@
       "start_time": 7.5,
       "step_description": "fold shorts horizontally",
       "step_id": 2,
-      "step_name": "fold shorts horizontally",
-      "step_unique_id": 88777554658133
+      "step_name": "fold shorts widthwise",
+      "step_unique_id": 106564956343095
     },
     {
       "best_view": "ego",
-      "end_time": 12.3,
+      "end_time": 14.067,
       "is_essential": true,
       "keystep_timing_source": "human",
       "start_time": 11.5,
       "step_description": "press shorts into shape",
       "step_id": 3,
-      "step_name": "press shorts into shape",
-      "step_unique_id": 159170461428859
-    },
-    {
-      "best_view": "ego",
-      "end_time": 14.067,
-      "is_essential": true,
-      "keystep_timing_source": "human",
-      "start_time": 12.3,
-      "step_description": "leave folded shorts on surface",
-      "step_id": 4,
-      "step_name": "leave folded shorts on surface",
-      "step_unique_id": 22536988523399
+      "step_name": "flatten shorts on bed",
+      "step_unique_id": 72292095753985
     }
   ],
   "object_segmentation": {

episodes/019fbdf4-04e8-7030-9358-724279af9698/depth_geometry.json

@@ -1,14 +1,18 @@
 {
   "sensor": "iphone_lidar_arkit",
-  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth, its OWN intrinsics \u2014 see depth_geometry.json); observation.images.ego is 256x256 (resized). DIFFERENT pixel grids \u2014 register via the depth intrinsics + observation.camera_pose, never depth[y,x]==ego[y,x]. The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
+  "note": "iPhone Pro LiDAR via ARKit sceneDepth \u2014 NOT a RealSense / structured-light depth cam (sparser + smoothed). observation.images.depth is 256x192 (native ARKit sceneDepth grid, untouched); observation.images.ego is 512x384 \u2014 ONE uniform downscale of the native RGB, so ego is EXACTLY 2x the depth grid: ego[y, x], mask[y, x] and depth[y//2, x//2] are the same ray. Intrinsics for both grids are in depth_geometry.json; camera_pose is expressed in the SHIPPED raster frame (episodes flagged raster_rotation=ccw90 were rotated from landscape \u2014 rotate 90\u00b0 CW to display upright). The lossless raw depth (metres + per-pixel confidence) rides in the OpenEgo sidecar depth.npz.",
   "depth_resolution_hw": [
     256,
     192
   ],
   "ego_obs_resolution_hw": [
-    256,
-    256
+    512,
+    384
   ],
+  "ego_to_depth_scale": 2.0,
+  "raster_rotation": null,
+  "display_hint": null,
+  "camera_pose_frame": "shipped_raster (sensor-frame poses corrected by 1 quarter-turn(s); gravity-verified at export)",
   "rgb_native_resolution_hw": [
     1920,
     1440
@@ -32,6 +36,7 @@
       1.0
     ]
   ],
+  "intrinsics_frame": "native raster",
   "depth_intrinsics_3x3": [
     [
       182.71277669270833,

episodes/019fbdf4-04e8-7030-9358-724279af9698/openego_sidecar.json

@@ -396,100 +396,100 @@
       },
       "quality_gates": {
         "all_passed": true,
+        "validated_at": "2026-08-10T13:35:49+00:00",
         "gates": [
           {
-            "detail": "",
             "name": "json_schema",
             "passed": true,
+            "value": "ok",
             "threshold": "validates",
-            "value": "ok"
+            "detail": ""
           },
           {
-            "detail": "presence-conditioned over the actor's active span [47..1031] = 985/1115 frames (raw all-frames = 82.78%)",
             "name": "frames_with_hands_pct",
             "passed": true,
+            "value": 100.0,
             "threshold": ">= 30% [class=GLOBAL]",
-            "value": 93.71
+            "detail": "presence-conditioned over the actor's active span [39..1038] = 1000/1115 frames (raw all-frames = 89.69%); shipped channel = raw \u2295 hand_mask_corrections"
           },
           {
-            "detail": "silent or near-silent",
             "name": "audio_snr_db",
             "passed": true,
+            "value": null,
             "threshold": ">= 5 dB",
-            "value": null
+            "detail": "silent or near-silent"
           },
           {
-            "detail": "silent clip",
             "name": "transcript_density",
             "passed": true,
+            "value": 0.0,
             "threshold": ">= 0.5 wps",
-            "value": 0.0
+            "detail": "silent clip"
           },
           {
-            "detail": "",
             "name": "hand_mask_coverage_pct",
             "passed": true,
+            "value": 2.354,
             "threshold": ">= 1%",
-            "value": 2.354
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "phase_plan_windowing",
             "passed": true,
+            "value": "not_requested",
             "threshold": "phase plan absent \u2014 skipped",
-            "value": "not_requested"
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "object_mask_coverage_pct",
             "passed": true,
+            "value": 16.77,
             "threshold": ">= 1% (per present object) [class=GLOBAL]",
-            "value": 16.77
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_left",
             "passed": true,
+            "value": 2.4022,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 2.4022
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_area_floor_right",
             "passed": true,
+            "value": 2.2951,
             "threshold": ">= 0.20% (median, per hand)",
-            "value": 2.2951
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_mask_lr_asymmetry",
             "passed": true,
+            "value": 0.9554,
             "threshold": "min/max >= 0.20",
-            "value": 0.9554
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "hand_slot_merge",
             "passed": true,
+            "value": 0,
             "threshold": "< 5 consecutive frames with L\u2229R IoU > 0.5",
-            "value": 0
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "object_track_truncated",
             "passed": true,
+            "value": 0,
             "threshold": "no tracked object ends while a hand still holds it",
-            "value": 0
+            "detail": ""
           },
           {
-            "detail": "",
             "name": "actor_mirror_consistency",
             "passed": true,
+            "value": 0,
             "threshold": "every hierarchy span's actor_kind == its action's actors[0]",
-            "value": 0
+            "detail": ""
           }
-        ],
-        "validated_at": "2026-08-04T08:55:45+00:00"
+        ]
       }
     },
     "manually_annotated_fields": [],