Dataset · Egocentric video
Egocentric video dataset
First-person video of hands, tools and tasks filmed from the worker’s point of view, licensed for training robotics, embodied AI and video models.
4.2M hours
of egocentric video available today
What’s in it
Real work recorded from the person doing it: the grip, the tool change, the mistake and the recovery. The parts of a task that finished-product footage never shows.
- Hands, tools and objects in frame
- Full tasks, start to finish
- Indoor, outdoor, industrial and household settings
What teams use it for
Pretraining and fine-tuning models that need to understand how physical work is done.
- Robot manipulation and imitation learning
- Action recognition and video understanding
- World models and embodied agents
How licensing works
01
Tell us the use
The model, the task and the volume you need. An NDA covers everything that follows.
02
Get a sample
We confirm the exact subset, rights and delivery format, and send a sample before anything is priced.
03
License
Every record ships scrubbed of personal data, with consent records and a written license.
04
Fill the gaps
If the inventory doesn’t cover it, Harbor Collect records it to your spec.
Questions
- Can you record a specific task we don’t see in the inventory?
- Yes. Harbor Collect commissions new first-person recordings from contributors who do the work, with review criteria set before collection begins.
- Is the data cleared for commercial model training?
- Rights are confirmed per project before pricing, and every delivery comes with a written license covering the agreed use.
- How is personal data handled?
- Faces, screens, spoken personal details and file metadata are scrubbed before delivery, and each redaction is logged in a provenance report.
More datasets
Training data for robot learning
Robots learn manipulation from watching the doing, not the done. Harbor supplies first-person video of real tasks and records new ones to your spec.
Conversational audio dataset
Natural, multi-speaker conversation at scale: call-centre recordings, podcasts, interviews and radio, licensed for speech and voice models.
Speech recognition training data
ASR models fail on real calls because they were trained on clean reads. Harbor licenses real conversation, noise and all.
CT, MRI and X-ray images with reports
Harbor’s Scientific Media collection pairs every image with its radiology report, ready for vision-language training.
Video dataset for AI training
Licensed footage for video generation and understanding, from street scenes to long-form conversation.
Egocentric video dataset
First-person video of hands, tools and tasks filmed from the worker’s point of view, licensed for training robotics, embodied AI and video models.