AgentVQA/AgentVQA_old

AgentVQA: A Multi-Domain Visual Question Answering Dataset

AgentVQA is a comprehensive dataset for training and evaluating visual agents across multiple domains, including GUI interaction, spatial reasoning, game understanding, robot manipulation, and video perception. The dataset contains multiple-choice questions based on screenshots, images, and videos.

Dataset Structure

The dataset is organized into five main domains:

1. Web Agents Domain (GUI Interaction)

• android-in-the-wild: Android app interactions (1000 samples)
• mind2web: Web navigation tasks (1000 samples, stratified by category)
• monday: Mobile OS interactions (1000 samples, stratified by OS)
• screenspot: GUI element grounding (1000 samples)
• screenspot-pro: Advanced GUI grounding (1000 samples)

2. Spatial Reasoning Domain

• embspatial-bench: Object-object and object-scene spatial relationships (1000 samples)
• space-10: Multi-image spatial reasoning with entity presence tasks (1000 samples)
• 3dsrbench: 3D spatial reasoning benchmark
• spatialmm: Spatial reasoning in multi-modal contexts
• omnispatial: Comprehensive spatial reasoning across multiple task types

3. Game Understanding Domain

• gameqa: 3D spatial perception and reasoning in games (1000 samples)
• videogamebunny: Video game understanding across multiple categories
• but-they-are-cats-tutorial: Game tutorial comprehension tasks
• marioqa: Video-based event understanding in Super Mario gameplay

4. Robot Manipulation Domain

• robo2vlm: Robot task success state evaluation (1000 samples)
• roborefit: Robot grasping and manipulation tasks (1000 samples)
• erqa: Embodied reasoning for robot question answering (400 samples)
• manipbench: Robot manipulation trajectory planning (1000 samples)
• sharerobot: Multi-frame robot task understanding (1000 samples)

5. Video Perception & Egocentric Understanding Domain

• perception-test: Camera motion and video perception tasks (1000 samples)
• vsi-bench: Video spatial intelligence with egocentric viewpoint questions (1000 samples)
• openeqa: Open-ended embodied question answering from video (1000 samples)
• videgothink: Egocentric video understanding and reasoning (1000 samples)
• egoplan-bench: Egocentric task planning and goal understanding (1000 samples)

Loading the Dataset

Load a specific dataset:

from datasets import load_dataset

# Load a web agents dataset
ds = load_dataset("advaitgupta/AgentVQA", "android-in-the-wild", split="train")

# Load a spatial reasoning dataset
ds = load_dataset("advaitgupta/AgentVQA", "embspatial-bench", split="train")

# Load a video perception dataset
ds = load_dataset("advaitgupta/AgentVQA", "perception-test", split="train")

# Load an egocentric dataset
ds = load_dataset("advaitgupta/AgentVQA", "egoplan-bench", split="train")

Loading Video Datasets (stored as bytes):

from datasets import load_dataset
from decord import VideoReader
from io import BytesIO

# Load video dataset
ds = load_dataset("advaitgupta/AgentVQA", "videgothink", split="train")
sample = ds[0]

# Access video bytes
video_bytes = sample["Video_Bytes"]
filename = sample["Video_Filename"]

# Option 1: Save to file
with open(f"/tmp/{{filename}}", "wb") as f:
    f.write(video_bytes)

# Option 2: Use decord for direct loading
vr = VideoReader(BytesIO(video_bytes))
frames = vr.get_batch(range(len(vr))).asnumpy()

Helper Functions:

from datasets import load_dataset

# Available domains and their datasets
DOMAINS = {{
    "web_agents": [
        "android-in-the-wild",
        "mind2web",
        "monday",
        "screenspot",
        "screenspot-pro"
    ],
    "spatial_reasoning": [
        "embspatial-bench",
        "space-10",
        "3dsrbench",
        "spatialmm",
        "omnispatial"
    ],
    "game_understanding": [
        "gameqa",
        "videogamebunny",
        "but-they-are-cats-tutorial",
        "marioqa"
    ],
    "robot_manipulation": [
        "robo2vlm",
        "roborefit",
        "erqa",
        "manipbench",
        "sharerobot"
    ],
    "video_perception": [
        "perception-test",
        "vsi-bench",
        "openeqa",
        "videgothink",
        "egoplan-bench"
    ]
}}

def load_domain_dataset(domain, dataset_name):
    \"\"\"Load a specific dataset from a domain.\"\"\"
    return load_dataset("advaitgupta/AgentVQA", dataset_name, split="train")

def load_all_domain_datasets(domain):
    \"\"\"Load all datasets from a specific domain.\"\"\"
    datasets = {{}}
    for dataset_name in DOMAINS[domain]:
        datasets[dataset_name] = load_dataset("advaitgupta/AgentVQA", dataset_name, split="train")
    return datasets

def load_all_datasets():
    \"\"\"Load all datasets from all domains.\"\"\"
    all_datasets = {{}}
    for domain in DOMAINS:
        all_datasets.update(load_all_domain_datasets(domain))
    return all_datasets

# Example usage:
ds = load_domain_dataset("video_perception", "perception-test")
video_datasets = load_all_domain_datasets("video_perception")
print(video_datasets.keys())

Dataset Fields

Web Agents Domain

android-in-the-wild

• image: Screenshot of the app
• step_id: Step number in the episode
• action_history: Previous actions taken
• question: The question about next action
• options: List of 4 possible actions
• correct_answer: Index of correct option (0-3)
• episode_goal: Overall goal of the episode
• source_dataset: Original source dataset name

mind2web

• image: Screenshot of the webpage
• question: The question about next action
• options: List of possible actions
• answer: Index of correct option
• task: Description of the task
• action_history: Previous actions taken
• category: Task category
• website: Website name
• domain: Domain category

monday

• image: Screenshot of the mobile interface
• goal: Goal description
• question: The question about next action
• options: List of possible actions
• answer: Index of correct option
• action_history: Previous actions taken
• os: Operating system (Android/iOS)

screenspot & screenspot-pro

• image: Screenshot
• question: Instruction to complete
• options: List of tap/click actions
• answer: Index of correct option

Spatial Reasoning Domain

embspatial-bench

• Image: Scene image
• Question: Spatial relationship question
• Options: List of 4 possible answers
• Answer: Index of correct option (0-3)
• Category: High-level category

space-10

• Images: List of 8 scene images (multi-image reasoning)
• Question: Entity presence question
• Options: List of 4 descriptions
• Answer: Index of correct option (0-3)
• Category: Task category

3dsrbench, spatialmm, omnispatial

• Image: Scene image
• Question: Spatial reasoning question
• Options: List of possible answers
• Answer: Index of correct option
• Category/Task_Type: Task categorization

Game Understanding Domain

gameqa

• Image: Game screenshot
• Question: Game state reasoning question
• Options: List of 4 possible answers
• Answer: Index of correct option (0-3)
• Game_Type: Type of game
• Game_Name: Specific game name

videogamebunny

• Image: Game screenshot
• Question: Game understanding question
• Options: List of 4 possible answers
• Answer: Index of correct option (0-3)
• Category: Question category

marioqa

• Video: Super Mario gameplay video clip
• Question: Event-based question about gameplay
• Options: List of 4 possible answers
• Answer: Correct option letter (A, B, C, or D)
• Question_Type: Type of question
• Event: Event type being queried

Robot Manipulation Domain

robo2vlm

• Image: Robot scene image
• Question: Task success evaluation question
• Options: List of 4 possible answers
• Answer: Index of correct option (0-3)
• Category: Question category

roborefit

• Image: Robot manipulation scene
• Question: Grasping instruction
• Options: List of 4 coordinate pairs
• Answer: Index of correct option (0-3)

erqa

• Images: List of robot scene images
• Question: Embodied reasoning question
• Options: List of 4 possible answers
• Answer: Index of correct option (0-3)
• Category: Question type

manipbench

• Image: Robot manipulation scene with grid overlay
• Question: Trajectory planning question
• Options: List of 4 trajectory options
• Answer: Index of correct option (0-3)
• Category: Task category

sharerobot

• Images: List of robot trajectory frames
• Question: Task understanding question
• Options: List of 4 possible task descriptions
• Answer: Index of correct option (0-3)
• Category: Task type

Video Perception & Egocentric Domain

perception-test

• Video_Bytes: Raw video bytes (decode with decord)
• Video_Filename: Original video filename
• Question: Perception question (e.g., camera motion)
• Options: List of possible answers
• Answer: Index of correct option (0-3)

vsi-bench

• Video_Bytes: Raw video bytes
• Video_Filename: Original video filename
• Question: Spatial intelligence question from egocentric viewpoint
• Options: List of possible answers (directional/spatial)
• Answer: Index of correct option (0-3)
• Category: Question type (e.g., object_rel_direction)

openeqa

• Video_Bytes: Raw video bytes
• Video_Filename: Original video filename
• Question: Open-ended question about the environment
• Options: List of possible answers
• Answer: Index of correct option (0-3)
• Category: Question category (e.g., attribute recognition, world knowledge)

videgothink

• Video_Bytes: Raw video bytes
• Video_Filename: Original video filename
• Question: Understanding question about egocentric actions
• Options: List of possible answers
• Answer: Index of correct option (0-3)

egoplan-bench

• Video_Bytes: Raw video bytes
• Video_Filename: Original video filename
• Question: Planning question about task execution
• Options: List of possible next actions
• Answer: Index of correct option (0-3)
• Task_Goal: Overall goal of the task

Citation

If you use this dataset, please cite the original papers for each component dataset.

License

MIT License