GCRL / General15 assets

面向 general goal pose conditioned low-level robot manipulation policy 的仿真资产库与 G01–G15 任务规范。

This release contains selected simulation object assets, contact-structure annotations, collision representations, validation evidence and a complete gallery. It does not contain completed General15 RL environments or policy-success results.

Release: general15-v2 — 1262 asset instances, 743 conservative independent budget units, 62 structure families. Parameter variants sharing a template do not count as independent structures.

详细任务及使用说明 / General15 README · 代码 / Code · 发布文件 / Release files

Download

从代码仓库根目录运行：

python -m pip install huggingface_hub
python -m general15.download

The installer downloads the archive, verifies SHA256, checks archive paths, and relocates runtime metadata to the local installation. Use --root /data/general15 for another destination, and --revision <HF commit SHA> to pin a revision. Existing assets are not overwritten unless --overwrite is explicitly supplied.

Manual download: general15-v2.tar.gz, release.json, file hashes, SHA256SUMS.

The archive expands to general15/{assets,data,reports,gallery,config,docs,snapshots} plus documentation. It contains only manifest-selected asset directories; unused expansion candidates and raw upstream libraries are excluded. Runtime paths in the archive are relative to general15/; external absolute paths are retained solely as source provenance.

Task taxonomy

ID	Task
G01	抓取、搬运与悬停 / Grasp, transport & hold
G02	支撑放置与倚靠 / Supported placement & leaning
G03	容器放入与取出 / Put into & take out
G04	精密插入与拔出 / Insert & extract
G05	套挂与解挂 / Mount, hang & disengage
G06	有限行程关节操作 / Bounded articulated motion
G07	连续旋转与螺旋操作 / Continuous rotation & screwing
G08	平面非抓取移动 / Planar nonprehensile manipulation
G09	滑到边缘再抓起 / Slide to edge, then grasp
G10	立入支架再换抓 / Stand in a holder, then regrasp
G11	工具勾拉、推动与传递扭矩 / Tool-mediated manipulation
G12	杠杆撬动 / Lever manipulation
G13	环境辅助重定向 / Environment-assisted reorientation
G14	手内重定位 / In-hand repositioning
G15	托举与非夹持搬运 / Nonprehensile support & transport

The full README defines process constraints, goal frames, success relations, category boundaries and current validation limits for every task. Task labels are overlapping candidates, not certificates of robot-task feasibility.

Inventory and validation

Metric	Count
Instances	1262
Conservative independent budget units	743
Structural label families	62
Parameter instances	324
Bullet asset QA pass / needs repair	1214 / 48
Isaac Gym import/body/DOF checks passed	1262
SDF fixed-actor execution checks passed	109
Train / validation / test	1014 / 124 / 124

Main roles: 519 rigid objects, 213 containers/supports/fixtures, 203 articulated objects, 120 tools, 207 assembly components. Sources: GSO (220), PhysX-Mobility (438), Infinigen (200), AnyPlace (100), AutoMate (117), Factory (44), LIBERO (16), and project-generated geometry (127).

Use data/isaac_imported_bullet_ready.json as a basic QA subset, then filter for the task, source split and tested pair relation. The complete manifest.json includes 48 unresolved assets. The 402 pair records distinguish 144 released-stability passes, 30 path-only passes, 213 failed/alternate-path cases and 15 structure-only candidates. A path-only pass may have failed release stability.

109 SDF adapters passed actor creation and 120 simulation steps with a fixed asset and dynamic sphere on GPU PhysX. This does not validate fine apertures, threads, dynamic object manipulation or robot tasks. One asset uses resolution 128 after a resolution-256 native cooking failure; the prior failure remains in the evidence. All other adapters use 256.

The repair round added 75 deduplicated instances from 29 priority templates and recovered 2 previously failed source models. Of the original 168 flagged assets, 120 now pass current QA, including actual collision repairs and separately recorded annotation/probe corrections. Only 14 are accepted refined collision decompositions. The remaining 48 failures comprise 25 joint-sweep collisions, 21 aperture/cavity failures and 2 settling failures.

Original-model, near-duplicate, parameter-template and known mating families remain grouped in fixed splits. Do not randomly split scaled or textured variants across train/test. Mass, friction, inertia and centre of mass are estimates/assumptions where marked, not measured ground truth. No assets were excluded because a learned policy failed.

Gallery

After download:

python -m http.server 8877 --bind 127.0.0.1 --directory general15

Forward port 8877 in VS Code Remote SSH and visit http://localhost:8877/gallery/index.html. All 1262 assets have actual mesh views; the gallery supports task, role, source, split, QA status and revision-group filters. This dataset repository stores the gallery inside the archive rather than hosting an interactive web application.

Source terms and attribution

This collection has mixed upstream terms, not a blanket MIT asset license. The previous empty dataset card's license: mit was insufficient for the combined assets. The code's license does not override object-source licenses.

Per-asset license and provenance preserve available source attribution: GSO and Infinigen record CC BY 4.0; PhysX-Mobility records CC BY-NC 4.0; Factory records IsaacGymEnvs BSD-3-Clause; AutoMate refers to NVIDIA asset terms. AnyPlace and LIBERO per-mesh licensing information is incomplete in the current inventory. Project-generated geometry is separately marked. See the full README and each asset record for the available information and original source URLs.

For reproducibility, cite the upstream datasets you use and record the HF revision, archive SHA256, selected IDs, collision representation and task-specific validation conditions.