File size: 3,452 Bytes
eaaf38d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | ---
license: cc-by-4.0
task_categories: [robotics]
tags: [robotics, manipulation, grasping, grasp-annotation, franka, maniguard]
pretty_name: ManiGuard Grasp Annotations
size_categories: [1K<n<10K]
---
# ManiGuard Grasp Annotations
The curated grasp database behind ManiGuard's automated demonstration generation:
**1,547 annotated 6-DoF grasps over 221 object instances** across the six
ManiGuard-Bench task families. Each grasp is an eef-link target pose in the object's
local frame for the long-finger Franka Panda gripper
([`franka-panda-longfinger`](https://huggingface.co/datasets/IDEAS-Lab-Northwestern/franka-panda-longfinger));
at runtime the world grasp target is simply `T_eef_world = T_object_world @ T_grasp_local`.
Grasps were placed and reviewed per object instance in a viser-based annotation tool
(`maniguard/data/datagen/annotation/` in the [codebase](https://github.com/NU-IDEAS-Lab/ManiGuard));
every object carries several alternatives (median 5, up to 40) so the motion planner can
fall back when a grasp is unreachable or blocked.
## Files
| File | Contents |
|---|---|
| `grasp_annotations.json` | the database: `objects` keyed by `category/model`, each with `bbox_size`, `upright_orientation_xyzw`, and a `grasps` list (`position`, `orientation_xyzw` in the object-local frame, `approach_hint`, `source`) |
| `cabinet_geom.json` | cached drawer-link geometry scalars for the cabinet family (slide axis/sign, stroke, root-local AABBs) |
`source` per grasp: `freedrag` (953) and `click` (582) — the tool's two manual placement
modes — plus `topdown_gen` (12) tool-proposed top-down poses.
## No upstream geometry
This dataset contains **poses and metadata only** — no meshes or other geometry derived
from the [BEHAVIOR-1K](https://github.com/StanfordVL/BEHAVIOR-1K) assets, per the
upstream asset terms. The `mesh` fields are relative path strings used by the annotation
tool. Running ManiGuard's datagen does **not** require meshes; to use the annotation
*tool* (add or edit grasps), regenerate them locally from your own BEHAVIOR-1K download
with `maniguard/data/datagen/annotation/extract_meshes.py`.
## Usage
ManiGuard's datagen loader (`maniguard.data.datagen.grasp_db`) reads the database from
`outputs/grasp_annotation/` under the repo root:
```bash
hf download IDEAS-Lab-Northwestern/maniguard-grasp-annotations \
--repo-type dataset --local-dir outputs/grasp_annotation
```
See the [data-collection docs](https://nu-ideas-lab.github.io/ManiGuard/data_collection/)
for the full scripted-datagen pipeline.
## Paper & Citation
Part of **ManiGuard** — [paper (arXiv:2608.17386)](https://arxiv.org/abs/2608.17386) ·
[code](https://github.com/NU-IDEAS-Lab/ManiGuard) ·
[docs](https://nu-ideas-lab.github.io/ManiGuard/)
```bibtex
@misc{peng2026maniguard,
title = {{MANIGUARD}: A Benchmark and Data Suite for Specification-Grounded
Safety Evaluation and Improvement of Robotic Manipulation},
author = {Peng, Yiyan and Wang, Philip and Zhan, Simon Sinong and Lyu, Yiqi
and Ni, Zhenyang and Yan, Jixin and Wong, Fiorelli and Jiao, Ruochen
and Yin, Hang and Cao, Xinyu and Shao, Huajie and Li, Manling
and Zhang, Ruohan and Zhu, Qi},
year = {2026},
eprint = {2608.17386},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2608.17386},
}
```
|