IndEgo: A Dataset of Industrial Scenarios and Collaborative Work for Egocentric Assistants

Vivek Chavan¹²*, Yasmina Imgrund²†, Tung Dao²†, Sanwantri Bai³†, Bosong Wang⁴†, Ze Lu⁵†, Oliver Heimann¹, Jörg Krüger¹²

¹Fraunhofer IPK, Berlin    ²Technical University of Berlin    ³University of Tübingen
⁴RWTH Aachen University    ⁵Leibniz University Hannover

^{*Project Lead     †Work done during student theses/projects at Fraunhofer IPK, Berlin.}

Published at NeurIPS 2025

---

📖 Abstract

We introduce IndEgo, a multimodal egocentric and exocentric video dataset capturing common industrial tasks such as assembly/disassembly, logistics and organisation, inspection and repair, and woodworking. The dataset includes 3,460 egocentric recordings (~197 hours) and 1,092 exocentric recordings (~97 hours).

A central focus of IndEgo is collaborative work, where two workers coordinate on cognitively and physically demanding tasks. The egocentric recordings include rich multimodal data — eye gaze, narration, sound, motion, and semi-dense point clouds.

We provide:

• Detailed annotations: actions, summaries, mistake labels, and narrations
• Processed outputs: eye gaze, hand poses, SLAM-based semi-dense point clouds
• Benchmarks: procedural/non-procedural task understanding, Mistake Detection, and reasoning-based Video QA

Baseline evaluations show that IndEgo presents a challenge for state-of-the-art multimodal models.

---

🧩 Citation

If you use IndEgo in your research, please cite our NeurIPS 2025 paper:

@inproceedings{Chavan2025IndEgo,
  author    = {Vivek Chavan and Yasmina Imgrund and Tung Dao and Sanwantri Bai and Bosong Wang and Ze Lu and Oliver Heimann and J{\"o}rg Kr{\"u}ger},
  title     = {IndEgo: A Dataset of Industrial Scenarios and Collaborative Work for Egocentric Assistants},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track},
  year      = {2025},
  url       = {https://neurips.cc/virtual/2025/poster/121501}
}

Acknowledgments & Funding

This work is funded by the German Federal Ministry of Education and Research (BMBF) and the German Aerospace Center (DLR) under the KIKERP project (Grant No. 01IS22058C) in the KI-Familie program. We thank the Meta AI team and Reality Labs for the Project Aria initiative, including the research kit, the open-source tools and related services. The data collection for this study was carried out at the IWF research labs and the test field at TU Berlin. Lastly, we sincerely thank the student volunteers and workers who participated in the data collection process.