MA-EgoQA: Question Answering over Egocentric Videos from Multiple Embodied Agents

Project Page | Paper | GitHub

MA-EgoQA (Multi-Agent Egocentric Video Question Answering) is a benchmark designed to evaluate models on their ability to understand multiple long-horizon egocentric video streams simultaneously collected from embodied agents.

Built on the EgoLife dataset, it features 266 hours of multi-agent video where 6 people lived together for 7 days. The benchmark includes 1.7k questions that specifically require reasoning across more than two agents' observations.

Question Categories

The benchmark covers five distinct categories:

Category	Abbr.	Description
Social Interaction	SI	Localizing conversations and group behaviors across video streams
Task Coordination	TC	How agents divide roles and collaborate toward shared goals
Theory of Mind	ToM	Reasoning about agents' beliefs, intentions, and mental states
Temporal Reasoning	TR	Concurrency and ordering of events across agents' timelines
Environmental Interaction	EI	Tracking distributed object usage across agents

Sample Usage

You can download the dataset using the Hugging Face CLI as follows:

huggingface-cli download KangsanKim71/MA-EgoQA --local-dir data --repo-type dataset

Citation

@misc{kim2026maegoqa,
    title={MA-EgoQA: Question Answering over Egocentric Videos from Multiple Embodied Agents}, 
    author={Kangsan Kim and Yanlai Yang and Suji Kim and Woongyeong Yeo and Youngwan Lee and Mengye Ren and Sung Ju Hwang},
    year={2026},
    eprint={2603.09827},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2603.09827}, 
}