Datasets:

bhatvineet
/

bopask-train

license: mit
task_categories:
  - visual-question-answering
  - question-answering
language:
  - en
tags:
  - robotics
  - 6dof-pose
  - grasping
  - spatial-reasoning
  - trajectory
  - depth-estimation
  - bop-challenge
size_categories:
  - 10M<n<100M
pretty_name: BOPASK-Train

BOPASK-Train

BOPASK is a large-scale visual-question-answering dataset for robotic spatial understanding, built on top of four BOP-Challenge: HANDAL, HOPE, LineMOD, and YCB-V.

This release contains 32.68 M training QA pairs across 172 K unique RGB images, covering 8 question types.

Family	QA pairs	Unique images
handal	442,729	4,416
hope	18,323,007	82,229
linemod	13,165,385	61,883
ycbv	749,461	23,524
Total	32,680,582	172,052

Question-type breakdown (all families combined)

Question type	Count
trajectory	13,552,164
spatial_reasoning	8,888,686
depth_relative	7,256,308
pose	1,438,641
grasp	853,401
depth_absolute	446,658
object_rearrangement	157,095
camera (extrinsics)	87,629

Layout

bopask-train/
├── handal/
│   ├── bopask-handal-train.jsonl
│   ├── images/        # RGB frames (.png)
│   ├── depth_maps/    # aligned depth (.png, uint16 mm)
│   └── masks/         # per-object binary masks (.png)
├── hope/     (same structure)
├── linemod/  (same structure)
└── ycbv/     (same structure)

All paths inside each jsonl are family-relative:

images/<basename>.png
depth_maps/<basename>_depth.png
masks/<basename>_..._mask.png (comma-separated if multiple objects)

Record format (LLaVA-style)

Each line of a *-train.jsonl is a JSON object:

{
  "conversations": [
    {"from": "user", "value": "<image>\n<question>"},
    {"from": "gpt",  "value": "<answer>"}
  ],
  "images":  ["images/scene_000000_frame_000000.png"],
  "depths":  ["depth_maps/scene_000000_frame_000000_depth.png"],
  "masks":   "masks/scene_000000_frame_000000_obj_000018_mask.png",
  "question_type":    "pose",
  "question_subtype": "matrix",
  "object_id": 18
}

Field notes

conversations: a user/assistant turn pair. The user prompt starts with the <image> sentinel token used by many VLMs (e.g. LLaVA / Qwen-VL).
images, depths: lists of paths relative to the family folder.
masks: a single string. If multiple masks are relevant (e.g. pairwise trajectory, spatial_reasoning, object_rearrangement questions) they are comma-separated: "masks/...target..._mask.png,masks/...goal..._mask.png". Some rows have masks: null when the question does not target a specific object (e.g. camera extrinsics). Masks are optional for most downstream uses.
object_id: integer for single-object questions, or a "target,goal" string for pairwise ones. Absent for camera questions.
question_type: one of pose, grasp, camera, depth_absolute, depth_relative, spatial_reasoning, trajectory, object_rearrangement.
question_subtype: further specifies the answer format (e.g. matrix / quaternion / 2dbbox / 3dbbox for pose; 2d / 3d for trajectory; etc.).

Known caveats

A very small number of depth / mask files (≈0.004% of rows, mostly in LineMOD scenes 12 & 39) are absent because the originals were not recoverable. The JSONLs still reference them so you may want to handle FileNotFoundError gracefully in your loader.
masks are not strictly required for most VQA training setups; downstream users who only need RGB + depth + the conversations can safely ignore them.

Quick start

import json
from datasets import load_dataset

ds = load_dataset(
    "bhatvineet/bopask-train",
    data_files={"handal":  "handal/bopask-handal-train.jsonl",
                "hope":    "hope/bopask-hope-train.jsonl",
                "linemod": "linemod/bopask-linemod-train.jsonl",
                "ycbv":    "ycbv/bopask-ycbv-train.jsonl"},
)
print(ds)

Or streaming one family at a time:

import json

path = "handal/bopask-handal-train.jsonl"
with open(path) as f:
    for line in f:
        rec = json.loads(line)
        # rec["images"][0] is relative to the "handal/" folder
        ...

Citation

If you use this dataset, please cite the BOPASK paper and the underlying BOP-Challenge object-pose datasets (HANDAL, HOPE, LineMOD, YCB-V).

License

Released under the MIT License for the question-answer annotations. The underlying RGB, depth, and mask assets inherit the license of their source BOP-Challenge datasets.