aegean-ai/ar4-assembly-vqa

AR4 assembly VQA

Questions and answers about building the AR4 MK4 robot arm, written against its assembly manual. Each row names a page of the manual and a box on that page, and the image a question refers to is that box rendered from the PDF. The dataset stores no images. You download the manual from Annin Robotics and cut the images yourself.

The dataset was built for an assignment that fine-tunes a small vision-language model, SmolVLM-256M-Instruct, on one product's documentation.

Source manual

The source is the AR4 MK4 Robot Manual, version 1.1, by Annin Robotics (377 pages):

https://raw.githubusercontent.com/Annin-Robotics/downloads/main/2026/02/AR4-MK4-Robot-Manul-1.1.pdf
sha256 44d5d8d60a94078f28744d449008dd5a80ef52ce46a400607ebb02d3eb8a7fd1

Check the sha256 before you use the boxes. Another revision of the manual lays out its pages differently, and the boxes would cut the wrong regions.

Columns

Column	Meaning
id	Row id
unit	The image unit the question belongs to. Rows with the same unit share an image.
kind	step for an assembly step, doc for a parts-list page
page	Page number in the manual, starting at 1
bbox	The image box on that page, [x0, y0, x1, y1] in PDF points
question	The question
answer	A short answer: a number, a part name, a size, a tool or a phrase
qtype	Question type, listed below
subassembly	J1 to J6 for assembly steps, bom for the parts list

To render one image with PyMuPDF:

import pymupdf
from PIL import Image

doc = pymupdf.open("AR4-MK4-Robot-Manul-1.1.pdf")
pix = doc[row["page"] - 1].get_pixmap(clip=pymupdf.Rect(*row["bbox"]), dpi=110)
image = Image.frombytes("RGB", (pix.width, pix.height), pix.samples)

Questions

Step questions (kind = step) come from the assembly chapter, pages 51 to 241. Each of those pages holds two panels, a bench photo with its printed instruction beside it. The box covers the whole panel, so the instruction is part of the image. The question types are count, fastener, fastener_count, fastener_list, part, tool, spec, action and joint.

Documentation questions (kind = doc) come from the bill of materials, pages 16 to 50. Grid pages of labeled part photos give name questions ("What is the top left photo called?") and where questions ("Which photo position shows the J2 ARM?"). The fastener table on page 28 gives table_count questions about how many of each fastener the whole robot needs. The labeled photos and their captions also give spec, count, part, joint and action questions.

Some pairs were made by rules applied to the instruction text, for example the counts printed as "(6) M3x20 flat head screws". Claude Sonnet 5.5 (claude-sonnet-5-5) wrote the others. It saw each unit's text, not its image, and worked in five batches under the instructions in qa_prompt.md. Those instructions ask for three to five short questions per unit and tell the model to copy every answer from the text, never from outside knowledge. A pair was kept only when the panel's text contains every number in the answer and at least 60% of its other words. Nobody has checked the pairs one by one, so expect some that are wrong or oddly worded.

Split

Split	Step	Doc	Total
train	1,013	136	1,149
test	297	47	344

Consecutive assembly steps reuse near-identical photos, so a random split would leak test images into training. The split is by page instead. The assembly pages are cut into blocks of ten, and every fifth block is held out for test: pages 71 to 80, 121 to 130, 171 to 180 and 221 to 230. Training pages directly beside a test block were dropped, and so were training pairs whose panel contains a photo that also appears in a test panel. For the parts list, every fourth unit and every fourth fastener-table row are test.

License

The questions and answers derive from the AR4 manual, which Annin Robotics publishes under the BSD 3-Clause License. The dataset is released under the same license, and LICENSE carries Annin's copyright notice:

Copyright (c) 2026, Annin Robotics

Annin Robotics did not produce or endorse this dataset.