toloka/homer-v2

HomER v2: Home Egocentric Robotics Dataset

HomER v2 is a curated dataset of egocentric household activity videos designed for robotics, embodied AI, and video understanding research.

The dataset contains 765 first-person videos representing approximately 100 hours of real-world household activities collected from 420 unique participants across 49 countries spanning 5 continents. Each video is accompanied by structured metadata and a natural-language description of the performed task.

Dataset Contents

Each record contains:

• Unique video identifier

• Activity category

• Natural language task description

• Video preview URL

• Video download URL

• Video duration

• Additional metadata

Dataset Composition

The dataset contains 765 first-person videos representing approximately 100 hours of real-world household activities distributed across 10 activity categories.

Activity Category	Videos	Hours
Indoor Plant Activities	85	9.49
Cleaning	95	9.98
Dining Table Tasks	88	10.01
Personal Care & Bathroom Activities	86	10.03
Desk & Office Tasks	78	10.01
Laundry & Garment Care	75	9.94
Home Organization	69	9.99
Repair & Assembly	67	10.09
Food Preparation	66	10.06
Crafts & Fine Manipulation	56	9.99
Total	765	99.60

Dataset Characteristics

• Egocentric (first-person) recordings
• Approximately 100 hours of video data
• Data collected from 420 unique participants across 49 countries
• Geographic coverage spanning Africa, Asia, Europe, North America, and South America
• Natural language task descriptions for every video
• Real-world household environments and activities
• Balanced distribution across activity categories
• Diverse object manipulation and interaction scenarios
• Suitable for robotics, embodied AI, video understanding, and multimodal model evaluation

Intended Use

HomER v2 is intended for:

• Robotics perception research

• Embodied AI evaluation

• Video understanding benchmarks

• Activity recognition

• Human-object interaction analysis

• Video-language model evaluation

• Task understanding and reasoning

Data Format

The dataset is distributed as structured metadata with references to source videos. Video assets are accessible through the provided preview and download URLs.

Curated by

Bakyt Yrysov (Toloka)

Oleksii Moga (Toloka)

Attribution

This dataset is released under the CC BY 4.0 license.

If you use HomER v2, please provide attribution to Toloka and cite the dataset in publications, benchmarks, and derivative works.

Citation

@dataset{homer_v2_2026,

  title     = {HomER v2: Home Egocentric Robotics Dataset},

  author    = {Toloka},

  year      = {2026},

  publisher = {Hugging Face}

}