You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Fusion 360 Gallery Assembly data is released by Autodesk under a custom license that permits NON-COMMERCIAL RESEARCH USE ONLY. This mirror is redistributed with permission obtained from the rights holder. By requesting access you agree that (a) you will use the data only for non-commercial research, (b) you will pass this same non-commercial research restriction on to anyone you share the data or any derivative with, and (c) you will attribute the upstream dataset (https://github.com/AutodeskAILab/Fusion360GalleryDataset). Access is granted manually. Please state your name, affiliation, and intended use.
Log in or Sign Up to review the conditions and access this dataset content.
Fusion 360 Gallery — Assembly 子集 (a1.0.0) · 真实设计装配体
Autodesk 从 Fusion 360 Gallery 真人上传的设计里抽出的多零件装配体库。本库把它转成了团队统一 的 7 字段 SFT schema:8,251 行 = 8,251 个装配体(一行一个装配体,不是一个零件)。
🚫 这不是 benchmark。 这个子集没有官方任务、没有官方指标、没有官方基线、没有 leaderboard。 JoinABLe 论文(CVPR 2022,arXiv:2111.12772)里所有定量表格跑的都是另一个子集 Joint Data(j1.0.0);论文对本子集只给了一段 use-case 设想,唯一用到它的 multi-part assembly 演示是定性的、装配顺序还是人手工给的。官方仓库里与本子集相关的三个工具(下载 / 转 CAD / 转图) 一个模型都没有。想发数字,评测口径得自建。把它当资源池,不要当 benchmark。
📄 License —— 非商用研究,且这条限制必须继续传给下游
上游是 Autodesk 自定义的非商用研究许可,不是 CC / MIT / Apache 任何一种 (LICENSE.md)。
- 仅限非商用研究用途。 LICENSE 第 1 条原文:*"You may access, use, reproduce and modify the Dataset, in each case, only for non-commercial research purposes."*
- 这条限制随数据传递。 你从本库拿到的数据、以及你基于它做出的任何派生集,都只能用于 非商用研究,并且你必须把同样的限制明确告知你的下游使用者(LICENSE 3.2 / 3.3)。
- 再分发经权利方许可(permission obtained)。 LICENSE 第 2 条的默认口径是禁止整包再分发; 本库的镜像是在取得权利方许可的前提下进行的,许可范围以我们与权利方的沟通为准。 署名与原仓库链接一并保留:上游 https://github.com/AutodeskAILab/Fusion360GalleryDataset。
- 另注:LICENSE 第 4 条禁止把模型与创作者 / 个人身份重新关联;第 11 条把你雇主也绑进本许可 —— 受雇于营利实体的人请特别注意。
- 引用请用 JoinABLe 的 CVPR 2022 出处(arXiv abs 页 Comments 字段与代码仓 README 都写了 CVPR 2022;数据仓 README 里那份 2021 preprint bibtex 别照抄)。
一个 sample 是什么
一行 = 一个装配体目录(如 7780_6c885e81/),不是一个零件。官方规模口径
8,251 assemblies / 154,468 separate parts(docs/assembly.md L8 = 仓库 README L73 = 论文 A.1.3,
三处逐字一致)——单位是装配体。joints / contacts / holes 全是跨 body 的引用(body UUID +
face/edge index),拆成零件行会把引用打断,所以我们按装配体成行;要单件视角的人按 step 的键展开即可。
⚠ 154,468 数的是「唯一 body」还是「occurrence 展开后的实例」,官方没说清。 这两套口径确实会分叉
(官方样例 belt_clamp:bodies 字典 2 个唯一 body,但 properties.body_count = 5)。
所以本库两套都数:本次转换实测 唯一 body 合计 154,468、body 实例合计 185,102,
每行的 metadata.n_bodies_unique / n_body_instances 也都给了。我们不替官方选边。
🔴 坑 1:joint 里没有 axis_line,也没有 transform
很多二手描述会说 joint 带 axis_line / transform——数据里根本没有这两个键。真实字段是
geometry_or_origin_one / geometry_or_origin_two 下的
origin / primary_axis_vector / secondary_axis_vector / tertiary_axis_vector
geometry_type / key_point_type / entity_one / type
joint 顶层只有 angle / geometry_or_origin_one / geometry_or_origin_two / is_flipped /
joint_motion / name / occurrence_one / offset / parent_component / timeline_index / type。
"joint axis" 是 JoinABLe 论文里由 B-Rep entity 推导出来的概念,不是 schema 里的键名。
本库的 annot.joints[*].geometry_one/geometry_two 就是按真实字段抽的。
🔴 坑 2:occurrence_two 可以整个缺失
当 joint 的一侧是 root 组件下的 body(而不是某个 occurrence)时,joint 里没有
occurrence_two 这个键,改由 entity 上的 root_component 标识。官方样例 belt_clamp 里
4/4 个 joint 都是这样。docs L192 把它列进元素表,但 L184 的措辞是 "can contain the following
information" —— 可含,不是必含。直接写 joint["occurrence_two"] 会 KeyError。
本次转换实测:**494 个 joint 没有 occurrence_two**(占全部 joint 的 3.1%)。
每个 joint 都带 occurrence_two_present 布尔位。entity 上的归属字段至少有三种形态
(occurrence / root_component / 两个都没有),本次实测带 root_component 的 entity 有
518 个——官方 assembly_graph.py 只做了「有没有 occurrence」二分支,照抄会漏掉这一路。
🔴 坑 3:顶层键可能缺失,也可能存在但值为 null
论文 A.1.3 写了:contacts 处理失败发生在 13% 的装配体、joints **2%**(失败时置 null)。
官方代码两种都防了。本库每行的 metadata.sections_status 分别记 ok / absent(没这个键)/
null(有键值为 None)。本次转换实测:
| 段 | ok / absent / null |
|---|---|
joints |
absent=6642 / null=142 / ok=1467 |
as_built_joints |
absent=7774 / null=51 / ok=426 |
contacts |
absent=1607 / null=1061 / ok=5583 |
holes |
absent=2836 / ok=5415 |
另:论文说只有 20% 的设计定义了 joint;本次实测有 joint 的装配体 1,467 个 (17.8%),有 contacts 的 5,583 个,有 holes 的 5,415 个。
🔴 坑 4:face/edge 的 index 官方是索引 .smt 的,不是 .step
docs L17 说得很清楚:B-Rep 以 .smt 为 ground truth,.step 只是 *"an alternate neutral B-Rep
file format"*,而 index 是 "referenced by index from the JSON back to entities in the .smt
file"。**index 与 .step 面序的对应关系,官方从未声明过。 本库收的是 .step(离线唯一能读的),
所以拿 index 去点 .step 的第 k 个面之前,请自己先验一遍对齐。
本库给了一个免费的交叉校验量:mesh_index 列里每个 body 有
step_advanced_face(从 STEP 文本数 ADVANCED_FACE)与 obj_faces(从 .obj 的 g face N
分组标记数出来的面数)。本次实测 148,859 个 body 两者相等、5,609 个不等、0 个无法比对。
⚠ 个数相等只说明面的个数一致,不证明顺序一致,别过度解读。
🔴 坑 5:hole 标签是自动跑的,没有人工复核
docs L419:*"we use the Autodesk Shape Manager feature recognition tool to identify and label holes in each part."* —— 算法自动生成,全程没有人工复核环节的任何记述,官方也没有给出任何准确率数字。 「没公布」不等于「没问题」。当"真实件上的特征识别真值"用之前,请自己抽样质检。
本次转换实测 138,690 个 hole(官方从未公布 holes 总数,论文只给了直方图),类型分布前 10:
| hole type | 个数 |
|---|---|
RoundHoleWithThroughBottom |
109,280 |
RoundHoleWithBlindBottom |
20,942 |
CounterboreThroughHole |
2,902 |
CountersunkHoleWithThroughBottom |
2,336 |
TaperedRoundHoleWithThroughBottom |
1,241 |
RoundBlindHoleWithConicalBottom |
715 |
CountersunkHoleWithBlindBottom |
565 |
CounterboreBlindHole |
311 |
RoundBlindHoleWithTaperedBottom |
102 |
RoundBlindHoleWithSphericalBottom |
95 |
其它要知道的
- 单位 cm,角度 radians(docs L61)。
- occurrence 的
transform不是 4x4 数组,是坐标系字典{origin, x_axis, y_axis, z_axis}(docs L104);4x4 是官方assembly_graph.py的transform_to_matrix()现场拼的。 几何文件都是 body 局部坐标,拼整机要沿tree逐级相乘。 - contact 容差 0.1 mm(docs L244)。本次实测 140,091 个 contact 面对。
- joint 类型 7 种(docs L230-236)。本次实测分布:
| joint type | 个数 |
|---|---|
RigidJointType |
9,882 |
RevoluteJointType |
4,150 |
SliderJointType |
979 |
CylindricalJointType |
617 |
PlanarJointType |
99 |
BallJointType |
53 |
PinSlotJointType |
30 |
- 官方
AssemblyGraph的节点 id 是occurrence_uuid + "_" + body_uuid,只有 root 下的 body 直接用 body_uuid。 - docs 与真实文件可能有零星出入(例如 components 的
part_number与number两种写法)—— 这条我们没有逐字核实到官方出处,只是按「两个键都接」的写法做了防御。所以assembly_json列保留原文,annot只是派生视图,两者不一致时以原文为准。
划分(两个都收了)
| 列 | 来源 | 本次实测分布 |
|---|---|---|
metadata.split_official |
train_test.json,随机 80% train / 20% test |
test=1,651 / train=6,600 |
metadata.split_cross |
train_test_cross.json,与 Joint 子集 j1.0.0 的跨集划分 |
test=1,533 / train=6,129 / unassigned=589 |
train_test_cross 的用途是:用 train joint 训模型、拿 test assembly 测,保证两边不重叠。
⚠ 官方自己在同一段里警告:train-test-cross 不代表 assembly 数据的正常分布、可能有偏
(docs L481 原文 *"does not represent a normal distribution of the assembly data, and be may be
biased"*,笔误照录)。unassigned = 划分文件里没这个 id;unknown = 本次转换没拿到该划分文件。
官方从未公布两个划分的确切条数,上表是本次转换实测的。
收了什么 / 没收什么
| 文件 | 收? | 为什么 |
|---|---|---|
assembly.json |
✅ 原文逐字(assembly_json 列) |
它就是这个子集的全部标注内容 |
<uuid>.step |
✅ 原文逐字(step 列) |
离线唯一能读的 B-Rep 中性格式;本次收了 154,468 个 |
<uuid>.obj |
⚠ 只扫不存(mesh_index 列) |
mesh 是 146 GiB 体量的主因、且可从 B-Rep 再生成;但它的 g face N / g halfedge H edge E 分组是包里唯一离线可见的 face/edge index 证据,所以只留分组统计 |
<uuid>.png |
❌ | Fusion 自带缩略图不是受控渲染;image 一律为 null,渲染由团队后续统一处理 |
<uuid>.smt |
❌ | Autodesk Shape Manager 私有格式,我们这条链路上没有能读它的工具 |
整装配体级 assembly.* 几何 |
❌ | 各 body + transform 可再生成 |
Schema
统一 7 字段 + 我们加的 3 列:
query—— 我们框定的问法(contacts / joints / holes 三问)。再说一次:子集本身没有官方任务。image—— null(原始物是 B-Rep + JSON,无图像)。annot—— 派生标注视图(JSON):joints/as_built_joints/contacts/holes/tree/root/occurrences/components/bodies/counts/sections_status/properties。bulky 的bounding_box与point_on_entity不进 annot(原文列里有)。reasoning—— null(源数据无 CoT)。cate=B/task=T-B4(CAD & 3D Geometry Understanding,与mfcad_original/mfinstseg/deepcad同族)。metadata——assembly_id、两个 split、各段计数、sections_status、bodies_without_step, 以及上面那几个坑的逐条说明(notes)。- ➕
assembly_json—— assembly.json 原文。annot是我们的派生视图,一旦解析口径有错,原文还在。 - ➕
step——{body_uuid: STEP 原文}的 JSON 对象(装配体是多 body 的,单字符串装不下)。 - ➕
mesh_index——{body_uuid: {obj_faces, obj_face_max, obj_edges, obj_halfedges, step_advanced_face}},给下游做不依赖 OCC 的面数自洽检查(见坑 4)。
Load
from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/fusion360_assembly")
Gated — request access; granted manually. 非商用研究用途,且该限制须继续传递给下游。
复现
原始数据不入库,只入 URL 与字节数。整包 146.53 GiB,分 11 个 7z 分卷
(合计 20,226,501,350 字节 = 18.84 GiB,逐卷 HTTP HEAD 实测)。转换器一卷一卷流式处理:
下载一卷 → 用 py7zr 的 extract(factory=...) 把成员直接解到内存里、读完即弃(不落地解压)→
产出该卷的 parquet 分片 → 删掉这一卷 → 下一卷。
set HF_TOKEN=...
python fusion360_asm/convert_fusion360_asm.py --volumes 0-10
python fusion360_asm/convert_fusion360_asm.py --push
本次转换用时合计 13 分钟,产出分片合计 3.26 GB。
Provenance
- 数据:https://github.com/AutodeskAILab/Fusion360GalleryDataset (Assembly 子集 a1.0.0,11 个 7z 分卷,2021-11-30)
- 文档:
docs/assembly.md(484 行,本子集唯一的官方文档) - 论文:Willis, Jayaraman, Chu, Tian, Li, Grandi, Sanghi, Tran, Lambourne, Solar-Lezama, Matusik. JoinABLe: Learning Bottom-up Assembly of Parametric CAD Joints. CVPR 2022(arXiv:2111.12772 v2)。
- 许可:见上「License」一节。请引用上游并保留署名。
- Downloads last month
- 10