Task 79 · Consumer-product assembly step reordering
Assembly101 source-native Dataset Report
报告对象：批准的固定单视角 30-execution 子集 v1；不是 Assembly101 约 3.89 TB 完整视频发布。
FULL VERIFIED · APPROVED SUBSETPARTIAL OF FULL RELEASESource A · Assembly101CC BY-NC 4.0
Decision · CONDITIONAL GO
该子集提供可靠的 Class B 监督：publisher 给出装配阶段与 fine-grained action 的真实时间边界，Task 79 的正确顺序可在后续 demo/pipeline 层确定性派生。数据集本身不提供“乱序输入 → 正确顺序 target”视频对。
技术上推荐继续。 商业用途必须暂停：CC BY-NC 4.0 不允许未经额外授权的商业使用。
Measured verification
30 / 30完整视频解码通过
46.24 GBpublisher 视频字节
695完整官方 annotation files
1,074,514完整解码帧数
30 / 30unique executions / toys
18 / 6 / 6train / validation / test
15 / 15attempt-correction / normal
0损坏、尺寸错误、越界
Source portfolio
| Role | Contribution | Native contents | Verification | License | Raw S3 |
|---|
Source A
Assembly101 subset v1 | 固定 RGB v4 完整执行；真实装配/返工时间序列 | 30 MP4 + complete 695-file annotation tree | FULL VERIFIED for approved subset
30 full decodes + full S3 SHA-256 readback | CC BY-NC 4.0 | 738 objects / 46,420,515,568 bytes |
Source contract vs Task contract
Publisher-native source contract
- 完整 1920×1080、60 fps、fixed RGB
C10119_rgb.mp4。 - fine annotations：video、start/end frame、action/verb/noun、toy、split、view type。
- coarse annotations：assembly/disassembly sequence 与 action intervals。
- 所有 annotation frame number 使用 30 fps timebase；视频本体为 60 fps。
- 不存在 publisher-native shuffled/ordered video pair。
Later demo contract · not Raw
- Input：同一 execution 的 N 个 annotation-bounded atomic clips，确定性乱序。
- Target：相同 clips 按 publisher 时间顺序排列。
- Prompt：恢复物理有效的装配顺序。
- Pairing key：execution/view ID + ordered native segment IDs。
- Metrics：exact permutation、Kendall τ、pairwise precedence、normalized edit distance。
以上全部属于 derived demo/pipeline，尚未创建，也不在 Raw S3。
Publisher-native sample evidence · 2 / 2
以下代理仅展示官方视频中的 annotation-bounded 时间段。它们不是 Task 79 的 input/target pair，也未上传 Raw。
Sample A · normal assembly chain
- Source role
- Source A · Assembly101 subset v1
- Split / ID
- train ·
nusar-2021_action_both_9086-c03d_9086_user_id_2021-02-16_153418 - Original path
recordings/.../C10119_rgb.mp4- Native bounds
- frames 2820–3155 @ 30 fps; 94.000–105.167 s
- Native labels
- pick up / rotate / inspect chassis; pick up and position cabin; pick up / position / screw cabin
- Selection case
normal_or_low_attempt; 0 native attempt segments in its assembly interval- Display changes
- trim; 1920×1080→960×540; 60→30 fps; muted; H.264/yuv420p
- Bytes / SHA-256
- 754,235 ·
dc7836ba…996d4
Sample B · attempt/correction chain
- Source role
- Source A · Assembly101 subset v1
- Split / ID
- validation ·
nusar-2021_action_both_9015-c10c_9015_user_id_2021-02-02_165955 - Original path
recordings/.../C10119_rgb.mp4- Native bounds
- frames 5588–5944 @ 30 fps; 186.267–198.133 s
- Native labels
- repeated attempt to position water tank; inspect water tank; attempt to position chassis
- Selection case
attempt_correction; 55 native attempt segments across its assembly interval- Display changes
- trim; 1920×1080→960×540; 60→30 fps; muted; H.264/yuv420p
- Bytes / SHA-256
- 675,961 ·
74a80624…1ac7
Measured subset composition
| Measure | Measured result | Interpretation |
|---|
| Publisher payload | 725 files / 46,420,185,904 bytes | 30 videos + 695 annotations; excludes HF cache |
| Videos | 46,235,733,880 bytes; 456,889,400 min; 1,272,886,824.5 median; 5,912,514,780 max | Complete fixed-view recordings, not atomic clips |
| Media distribution | 30× H.264; 30× yuv420p progressive; 30× 1920×1080; 30× 60 fps | Measured independently for all 30 files |
| Assembly fine segments | 6,504 | Overlapping publisher action intervals within native assembly bounds |
| Attempt segments | 542 | Native verb begins with attempt |
| Correction-like coverage | 1,719 | Selection heuristic over native position/attach/screw/insert verbs; not a publisher score |
Native layout and annotation schema
task79_subset_v1/
├── publisher/
│ ├── annotations/ # complete official tree, 695 files
│ │ ├── fine-grained-annotations/
│ │ └── coarse-annotations/
│ └── recordings/<execution_id>/C10119_rgb.mp4 # 30 files
└── _provenance/ # selection/download/audit/readback evidence
Fine CSV 原生字段：id, video, start_frame, end_frame, action_id, verb_id, noun_id, action_cls, verb_cls, noun_cls, toy_id, toy_name, is_shared, is_RGB。Coarse label 原生字段为 30 fps 的 start_frame, end_frame, action_cls。官方 annotation README 声明 1,380 fine actions（90 objects × 24 verbs）和 202 coarse actions（61 objects × 11 verbs）。
Integrity and archive verification
| Gate | Result |
|---|
| Download scope | 30/30 selected videos; complete annotations; 0 failures |
| Full decode | 30/30; 1,074,514 decoded frames; 0 decode failures |
| Size and annotation bounds | 30/30 exact publisher bytes; 0 out-of-bounds assembly intervals |
| Diversity/leakage controls | 30 unique execution IDs; 30 unique toy IDs; 18/6/6 split preserved |
| S3 Raw upload | publisher 725 objects / 46,420,185,904 bytes |
| Full S3 readback | 725/725 files; exact bytes; complete SHA-256 manifest match |
Official claims vs measured scope
| Topic | Official / parent release | This report measured |
|---|
| Scale | Assembly101 complete multi-view release is approximately 3.89 TB; most sequences have 12 views (8 fixed RGB + 4 monochrome egocentric). | One fixed RGB v4 view from 30 deterministic executions, 46.24 GB videos. |
| Annotations | Fine action recognition/anticipation and coarse temporal segmentation annotations. | Complete 695-file annotation tree, 184,452,024 bytes. |
| Status | Full multi-view release not ingested. | FULL VERIFIED for approved subset; PARTIAL relative to parent release. |
Limitations and next gate
- License：CC BY-NC 4.0; research evaluation is supported, commercial use requires separate permission.
- Subset：30 deterministic executions provide tens-of-GB feasibility, not complete multi-view coverage.
- No native pair：shuffling, target ordering, prompts and permutation metadata are future Class B derivations.
- Annotation overlap：fine actions can overlap in time; demo builder must freeze an atomic-clip overlap/merge policy before generation.
- Benchmark separation：Assembly101 official benchmarks are action recognition, anticipation and temporal segmentation; Task 79 permutation metrics are proposed downstream metrics, not official Assembly101 scores.
Next: after human acceptance of this Raw report, build a separate traceable demo/pipeline. Do not back-port derived clips or prompts into Raw.