Update verified Table Clean correspondence tables and validation

Verify the full local dataset against remote content hashes; synchronize changed managed files and add per-frame Aria/robot correspondence, review windows, and visual checks. Preserve original robot state/action data and document temporal uncertainty.

.gitattributes

@@ -63,3 +63,4 @@ source_metadata/mps_complete/mps_796d9b1b-5171-4947-83d3-295fda70d48a_vrs/hand_t
 source_metadata/mps_complete/mps_a7fafb07-84a0-4afe-9799-2a2fd5da4cc8_vrs/hand_tracking/hand_tracking_results.csv filter=lfs diff=lfs merge=lfs -text
 source_metadata/mps_complete/mps_b1ea6dea-f91a-49b9-beac-48bc2381d08b_vrs/hand_tracking/hand_tracking_results.csv filter=lfs diff=lfs merge=lfs -text
 source_metadata/vrs/796d9b1b-5171-4947-83d3-295fda70d48a/mps/hand_tracking/hand_tracking_results.csv filter=lfs diff=lfs merge=lfs -text
+correspondence/frame_correspondence.csv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,19 @@
+# Table Clean — robot trajectories with Aria gaze and hands
+
+200 episodes · 43,727 frames · 4 tasks · 30 FPS · LeRobot v3.0.
+
+This version contains the verified robot state/action records, robot head and chest videos, Aria RGB video, eye gaze, and left/right hand tracking. Original source metadata and depth archives are retained in `source_metadata/`.
+
+## Correspondence
+
+`correspondence/frame_correspondence.csv` provides one row per robot frame. Join using `(episode_index, frame_index)` to locate the source robot record, VRS recording, Aria RGB index, and zero-based gaze/hand CSV rows (excluding the header). The table includes validity, hand confidence, and nearby review windows. See [the correspondence guide](correspondence/README.md), [recording-level evidence](correspondence/recording_sync_quality.csv), and [visual samples](correspondence/previews/).
+
+Pairing uses original robot host timestamps and each VRS device-to-realtime clock anchor. This is approximate temporal correspondence, not certified hardware synchronization. Independent checks compare the same Aria wearer's hand as seen by the robot camera and Aria MPS, plus 30 three-view visual checkpoints. Small timing offsets remain uncertain; review windows are candidates for inspection/modeling, not proven error bounds. No inferred delay between different people's gestures and robot reactions has been applied as a clock correction.
+
+Gaze is valid in 43,727 frames; left-hand tracking in 38,619; right-hand tracking in 43,659. Use the validity masks; missing hand coordinates are not real zero poses. Hand 3D coordinates remain in the Aria device frame. Spatial transformation to robot coordinates is not calibrated.
+
+## Validation
+
+[Dataset verification](processing/verification_report.json) checks all source timestamps, robot states/actions, 600 video frame counts, loading/decoding, and preserved depth archives. [Correspondence verification](correspondence/correspondence_report.json) covers all 43,727 frame mappings.
+
+RGB nearest-sample residual is at most 16.67 ms; this is not an absolute cross-device synchronization bound. Aria images are rotated 90 degrees clockwise and gaze pixels are rotated consistently; see `meta/image_orientation.json`.

correspondence/README.md

@@ -0,0 +1,30 @@
+# 机器人轨迹与 Aria gaze / hand 对应
+
+正式数据集：`/media/g1/MobileDisk/table_clean/lerobot`。
+
+200 段轨迹、43,727 帧均已建立对应关系。`frame_correspondence.csv` 每一行对应一个机器人帧，可用 `(episode_index, frame_index)` 与 LeRobot 数据连接。`source_frame_idx` 和 `source_json` 可追溯到原始机器人记录；`vrs`、`aria_rgb_index`、`gaze_csv_row_0based`、`hand_csv_row_0based` 指向原始眼镜数据。CSV 行号均从 0 开始且不包含表头。
+
+时间主线：`aria_device_ns = robot_host_ns - (VRS realtime_clock_sample_ns - monotonic_clock_sample_ns)`，再选取最近 RGB、gaze 和 hand 采样。原始机器人状态、动作和时间戳保留。眼镜视频在正式数据中顺时针旋转 90 度，gaze 像素随之旋转；三维 hand 仍是 Aria 设备坐标，未做机器人坐标外参标定。
+
+## 复核结果与使用方式
+
+- 已核验全部帧的原始索引、时间换算和 gaze/hand 原始行号。
+- `previews/index.html` 覆盖 5 条录制，每条前后各选一段，合计 30 个时刻，可比较同一人的动作、同一物体及机器人姿态。
+- 额外从机器人头部相机检测同一位眼镜佩戴者的右手，再与 Aria 手部数据比较。前 4 条录制的小偏移候选约 100–200 ms；最后一条检测样本较少，分组结果不稳定。它们是诊断估计，未强行写入正式数据的时钟。
+- 每帧附带 `gaze_window_start_row_0based` / `gaze_window_end_row_exclusive` 以及相应 hand 范围。前 4 条使用 ±250 ms、最后一条使用 ±500 ms 的**人工复核/建模候选窗口**；这不是已证明的误差上界。通常先用 nominal 行号；需要对时间不确定性更稳健时再使用窗口内数据。
+- `gaze_valid`、`left_hand_valid`、`right_hand_valid` 和手部 confidence 应一起使用。无效 hand 不作真实零坐标使用，也没有用插值伪造缺失手势。
+- 未启用硬件时钟同步。最近采样点的毫秒残差不能解释为绝对跨设备误差。
+
+## 文件
+
+- `frame_correspondence.csv`：43,727 行逐帧对应与窗口。
+- `recording_sync_quality.csv`：5 条录制的时间复核证据。
+- `correspondence_report.json`：最终对应检查报告。
+- `same_person_sync_check.json`：同人手部诊断曲线。
+- `previews/index.html`：三视角抽查页面。
+
+手部关键点格式参考 [Aria 官方说明](https://facebookresearch.github.io/projectaria_tools/docs/data_formats/mps/hand_tracking)，独立检测方法使用 [MediaPipe Hand Landmarker](https://developers.google.com/edge/mediapipe/solutions/vision/hand_landmarker/python)。
+
+## 在 Hugging Face 上使用
+
+`source_json_repo_path`、`gaze_repo_path`、`hand_repo_path` 是仓库内的相对路径，可直接用于下载相应源文件。原始绝对路径仅用于本地追溯。视频用 `(episode_index, frame_index)` 通过 LeRobot 元数据读取；`aria_rgb_index` 是原始 VRS 内的帧号，VRS 本体未包含在这个训练数据仓库中。

correspondence/correspondence_report.json

@@ -0,0 +1,158 @@
+{
+  "episodes": 200,
+  "frames": 43727,
+  "source_timestamp_frame_and_offset_checks_passed": true,
+  "residual_ms": {
+    "rgb_delta_ns": {
+      "median": 7.807045,
+      "p95": 15.727272299999994,
+      "max": 16.665447
+    },
+    "gaze_delta_ns": {
+      "median": 25.540685,
+      "p95": 47.196095,
+      "max": 49.9911
+    },
+    "hand_delta_ns": {
+      "median": 7.792715,
+      "p95": 15.7351275,
+      "max": 16.663717
+    }
+  },
+  "valid_frames": {
+    "gaze_valid": 43727,
+    "left_hand_valid": 38619,
+    "right_hand_valid": 43659
+  },
+  "recordings": [
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/a7fafb07-84a0-4afe-9799-2a2fd5da4cc8.vrs",
+      "episodes": 70,
+      "frames": 16572,
+      "motion_check": {}
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/b1ea6dea-f91a-49b9-beac-48bc2381d08b.vrs",
+      "episodes": 69,
+      "frames": 14382,
+      "motion_check": {}
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/5efd9659-56a2-4e00-8e04-8670a62a396e.vrs",
+      "episodes": 40,
+      "frames": 8393,
+      "motion_check": {}
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/796d9b1b-5171-4947-83d3-295fda70d48a.vrs",
+      "episodes": 15,
+      "frames": 3013,
+      "motion_check": {}
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/ece055d2-c57b-4a94-aeb3-61fa872a7d12.vrs",
+      "episodes": 6,
+      "frames": 1367,
+      "motion_check": {}
+    }
+  ],
+  "applied_additional_offset_ms": {
+    "/media/g1/MobileDisk/ARIA/Table_clean/a7fafb07-84a0-4afe-9799-2a2fd5da4cc8.vrs": 0,
+    "/media/g1/MobileDisk/ARIA/Table_clean/b1ea6dea-f91a-49b9-beac-48bc2381d08b.vrs": 0,
+    "/media/g1/MobileDisk/ARIA/Table_clean/5efd9659-56a2-4e00-8e04-8670a62a396e.vrs": 0,
+    "/media/g1/MobileDisk/ARIA/Table_clean/796d9b1b-5171-4947-83d3-295fda70d48a.vrs": 0,
+    "/media/g1/MobileDisk/ARIA/Table_clean/ece055d2-c57b-4a94-aeb3-61fa872a7d12.vrs": 0
+  },
+  "absolute_clock_sync_verified": false,
+  "motion_method": "VR operator and Aria wearer are different actors; VR-hand lag is NOT used for clock correction. See same_person_sync_check.json and previews for independent checks.",
+  "final_method": "Original robot timestamps + VRS clock tags + nearest MPS/RGB sample; independent same-person and 30 visual checkpoint audit; no inferred reaction lag applied",
+  "verified_at": "2026-09-14T19:55:54.105800+08:00",
+  "same_person_profiles": [
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/a7fafb07-84a0-4afe-9799-2a2fd5da4cc8.vrs",
+      "same_person_detected_samples": 1565,
+      "estimated_best_lag_ms": 100,
+      "split_best_lags_ms": [
+        150,
+        100
+      ],
+      "motion_score_nominal": 0.25383377313786437,
+      "motion_score_best": 0.4018153017919726,
+      "evidence_episodes": 70,
+      "review_window_halfwidth_ms": 250,
+      "window_is_error_bound": false,
+      "applied_extra_shift_ms": 0
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/b1ea6dea-f91a-49b9-beac-48bc2381d08b.vrs",
+      "same_person_detected_samples": 1184,
+      "estimated_best_lag_ms": 150,
+      "split_best_lags_ms": [
+        150,
+        150
+      ],
+      "motion_score_nominal": 0.24463895571172684,
+      "motion_score_best": 0.42597371566730213,
+      "evidence_episodes": 65,
+      "review_window_halfwidth_ms": 250,
+      "window_is_error_bound": false,
+      "applied_extra_shift_ms": 0
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/5efd9659-56a2-4e00-8e04-8670a62a396e.vrs",
+      "same_person_detected_samples": 796,
+      "estimated_best_lag_ms": 200,
+      "split_best_lags_ms": [
+        250,
+        200
+      ],
+      "motion_score_nominal": 0.10012799840634035,
+      "motion_score_best": 0.2287528275357596,
+      "evidence_episodes": 40,
+      "review_window_halfwidth_ms": 250,
+      "window_is_error_bound": false,
+      "applied_extra_shift_ms": 0
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/796d9b1b-5171-4947-83d3-295fda70d48a.vrs",
+      "same_person_detected_samples": 274,
+      "estimated_best_lag_ms": 200,
+      "split_best_lags_ms": [
+        200,
+        200
+      ],
+      "motion_score_nominal": 0.06454760468780119,
+      "motion_score_best": 0.25025851021638673,
+      "evidence_episodes": 14,
+      "review_window_halfwidth_ms": 250,
+      "window_is_error_bound": false,
+      "applied_extra_shift_ms": 0
+    },
+    {
+      "vrs": "/media/g1/MobileDisk/ARIA/Table_clean/ece055d2-c57b-4a94-aeb3-61fa872a7d12.vrs",
+      "same_person_detected_samples": 83,
+      "estimated_best_lag_ms": 400,
+      "split_best_lags_ms": [
+        100,
+        2850
+      ],
+      "motion_score_nominal": 0.30230127545268615,
+      "motion_score_best": 0.42570449353242434,
+      "evidence_episodes": 4,
+      "review_window_halfwidth_ms": 500,
+      "window_is_error_bound": false,
+      "applied_extra_shift_ms": 0
+    }
+  ],
+  "visual_checks": {
+    "recordings": 5,
+    "episodes": 10,
+    "moments": 30,
+    "result": "No obvious episode mismatch or multi-second phase mismatch at sampled points; not a frame-accurate synchronization guarantee"
+  },
+  "rejected_vr_operator_motion_lag": true,
+  "formal_dataset": "/media/g1/MobileDisk/table_clean/lerobot",
+  "formal_dataset_reverification_passed": true,
+  "formal_dataset_verification_report": "/media/g1/MobileDisk/table_clean/processing/verification_report.json",
+  "completed_at": "2026-09-14T19:57:39.313775+08:00"
+}

correspondence/frame_correspondence.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7904929d8e1e81590366ab2bf9aa1080a297b0391d59ab4ac2c3adf62f53ed9f
+size 39989976

correspondence/previews/index.html

@@ -0,0 +1 @@
+<!doctype html><meta charset="utf-8"><title>逐帧对应抽查</title><style>body{font-family:sans-serif;background:#171717;color:#eee;margin:24px}img{max-width:100%}a{color:#acf}</style><h1>机器人轨迹、眼镜视频与 gaze 对应抽查</h1><p>每条 VRS 选择前后两个 episode，每段取三个时刻，共 30 个检查点。左为机器人头部视角，中为机器人胸部视角，右为眼镜佩戴者视角。黄色十字为 gaze。请比较同一物体、同一个人的动作或同一机器人姿态；不同人的手势和机器人响应之间可以存在真实延迟。</p><p>画面中的毫秒数是最近采样点残差，不是绝对时钟同步保证。没有把人体手坐标冒充为机器人坐标。</p><h2>episode_0015 · transparent_tape</h2><img src="episode_0015.jpg"><h2>episode_0056 · black_tape</h2><img src="episode_0056.jpg"><h2>episode_0085 · blue_cotton_swab_container</h2><img src="episode_0085.jpg"><h2>episode_0125 · brown_cotton_swab_container</h2><img src="episode_0125.jpg"><h2>episode_0149 · blue_cotton_swab_container</h2><img src="episode_0149.jpg"><h2>episode_0172 · transparent_tape</h2><img src="episode_0172.jpg"><h2>episode_0184 · black_tape</h2><img src="episode_0184.jpg"><h2>episode_0192 · black_tape</h2><img src="episode_0192.jpg"><h2>episode_0197 · black_tape</h2><img src="episode_0197.jpg"><h2>episode_0200 · black_tape</h2><img src="episode_0200.jpg">

correspondence/recording_sync_quality.csv

@@ -0,0 +1,6 @@
+vrs,same_person_detected_samples,estimated_best_lag_ms,split_best_lags_ms,motion_score_nominal,motion_score_best,evidence_episodes,review_window_halfwidth_ms,window_is_error_bound,applied_extra_shift_ms
+/media/g1/MobileDisk/ARIA/Table_clean/a7fafb07-84a0-4afe-9799-2a2fd5da4cc8.vrs,1565,100,"[150, 100]",0.25383377313786437,0.4018153017919726,70,250,False,0
+/media/g1/MobileDisk/ARIA/Table_clean/b1ea6dea-f91a-49b9-beac-48bc2381d08b.vrs,1184,150,"[150, 150]",0.24463895571172684,0.42597371566730213,65,250,False,0
+/media/g1/MobileDisk/ARIA/Table_clean/5efd9659-56a2-4e00-8e04-8670a62a396e.vrs,796,200,"[250, 200]",0.10012799840634035,0.2287528275357596,40,250,False,0
+/media/g1/MobileDisk/ARIA/Table_clean/796d9b1b-5171-4947-83d3-295fda70d48a.vrs,274,200,"[200, 200]",0.06454760468780119,0.25025851021638673,14,250,False,0
+/media/g1/MobileDisk/ARIA/Table_clean/ece055d2-c57b-4a94-aeb3-61fa872a7d12.vrs,83,400,"[100, 2850]",0.30230127545268615,0.42570449353242434,4,500,False,0

correspondence/visual_review.json

@@ -0,0 +1,18 @@
+{
+  "checked_episode_ids": [
+    "episode_0015",
+    "episode_0056",
+    "episode_0085",
+    "episode_0125",
+    "episode_0149",
+    "episode_0172",
+    "episode_0184",
+    "episode_0192",
+    "episode_0197",
+    "episode_0200"
+  ],
+  "moments_per_episode": 3,
+  "method": "Visual comparison of same person posture, same robot posture, and object grasp/release state across robot head/chest and Aria views",
+  "finding": "Corresponding interaction phases visible at all 30 sampled times; this is coarse validation, not a certified frame-level time error bound",
+  "nominal_mapping_retained": true
+}