← Dashboard · Dataset Report

TASK 79 · TRACEABLE DERIVED V2V DEMO · 2026-07-21

消费品装配步骤乱序重排

两组 Demo 均来自已审计的 Assembly101 Raw,但与 Dataset Report 展示的 c03d / c10c 完全不同。Input 把同一条 recording 中的 5 个官方 fine-action clips 固定打乱;Target 只把这 5 个相同编码 clips 恢复到原时间顺序。

2 DISTINCT EXECUTIONSN = 5CLASS B · ANNOTATION-DERIVEDNO RAW MUTATION
2独立 Demo
10原子步骤
30 fps854×480 H.264
0生成式 Target

Demo 1 · Train · Cabin attachment

task79_000000 · toy b06a · nusar-2021_action_both_9016-b06a… · C10119_rgb.mp4

① INPUT · shuffled clips[4, 2, 5, 1, 3]
② TARGET · temporal order[1, 2, 3, 4, 5]
STEP 1 · 0049779position cabin
STEP 2 · 0049791pick up screw
STEP 3 · 0049803position screw on cabin
STEP 4 · 0049827screw cabin with screwdriver
STEP 5 · 0049839rotate partial toy
③ Prompt

Reorder the five atomic assembly clips into the correct chronological order. Output only the reordered video, preserving every clip exactly once.

④ Metadata

shuffle seed represented by fixed permutation
target key = recording + view + segment IDs

metadata.json

Demo 2 · Test · SUV roof → front wheel

task79_000001 · toy a27 (SUV) · nusar-2021_action_both_9046-a27… · C10119_rgb.mp4

① INPUT · shuffled clips[5, 2, 4, 1, 3]
② TARGET · temporal order[1, 2, 3, 4, 5]
STEP 1 · 0009831pick up roof
STEP 2 · 0009843position roof
STEP 3 · 0009891position screw on first wheel
STEP 4 · 0009915position first wheel
STEP 5 · 0009927screw first wheel with screwdriver
③ Prompt

Reorder the five atomic assembly clips into the correct chronological order. Output only the reordered video, preserving every clip exactly once.

④ Metadata

test split · same recording/view
publisher fine-action frame bounds @30fps

metadata.json

Demo / Raw boundary

Publisher-native:continuous source MP4、fine-action IDs、labels、frame bounds、recording/view keys。
Derived demo:原子切片、固定 permutation、Input/Target 拼接、Prompt、Metadata、浏览器转码。Assembly101 并未发布乱序/正序视频 pair;这些文件不属于 Raw。

Canonical provenance:task79_demo_manifest.json

Recommended evaluation:exact permutation accuracy、Kendall’s τ、pairwise precedence accuracy、normalized edit distance;按 N 和是否含 attempt/correction 干扰分别报告。